Claims
Cross-source engagement, derived from relations (supports / contradicts / refines / extends) that later extractions declare against earlier propositions. Most claims — 8416 of 10413 — are single-source: you've been exposed to the idea, but no other document in the corpus has corroborated, disputed, or built on it yet. That's expected, not a problem; this page surfaces the minority that other reading has actually engaged with.
Live controversies (461)
Propositions another document explicitly contradicts — the ⚡ disputed & corroborated ones (also supported by other reading) are grouped first, then the rest by how much cross-source engagement they've drawn.
Connection-level authorization alone is insufficient governance; organizations must still supply their own runtime, per-action controls for what an agent does once it is inside a system (defense-in-depth). ⚡ disputed & corroborated
The human does not leave the loop but moves up a level — from reading every diff to owning the parts that do not transfer to a model: accountability (a model cannot be paged at 3am), the judgment of whether a change is the right one to build, the high-blast-radius gates, and the behavior nobody specified — so 'human in the loop' becomes 'human on the loop', sampling and auditing rather than reading every PR. ⚡ disputed & corroborated
Mythos Preview can autonomously find and exploit zero-day vulnerabilities in every major operating system and every major web browser ⚡ disputed & corroborated
The gateway + registry pattern is the dominant architectural consensus for deploying MCP at enterprise scale, serving as the control plane for all agent interactions. ⚡ disputed & corroborated
contested byMCP Is Making API Sprawl Worse · p10
Exposure 3 — vendor lock-in: enterprise AI deployments are not truly modular but Jenga-like, with the model a load-bearing piece woven into application logic (every prompt, workflow, safety filter and evaluation metric calibrated to one provider), so AI lock-in is more acute than classic ERP/cloud lock-in, and with no interoperability/portability standard or regulatory floor the ability to simply switch providers is largely illusory. ⚡ disputed & corroborated
Multi-agent systems shift the coordination burden from the tool to the user, creating a new management problem rather than solving the original one ⚡ disputed & corroborated
Adding more context (context files, ADRs, rules) to guide agent behavior degrades output quality rather than improving it, because summarization quality degrades as context grows. ⚡ disputed & corroborated
AI is making developers busier, not more productive: a large-scale study reframes AI-assisted coding as a redistribution of effort in which task-level output soars while net shipped value barely moves. ⚡ disputed & corroborated
Team standards should be encoded as versioned, repository-resident AI instructions rather than wiki documentation or tribal knowledge, because an instruction's execution is its application — it does not depend on a developer remembering and applying it under time pressure — which makes governance a side effect of the normal workflow rather than a separate disciplined step. ⚡ disputed & corroborated
contested byAudit your Agent files · p4
Reading code is the mechanism that produces accountability for its quality, security, and correctness. ⚡ disputed & corroborated
contested byPodcast: The AI Joy Gap: Why Some Developers Thrive While Others Struggle · p9, What's an Orchestrator—and Why Does Software Need One? · p11
supported byWho cleans up after the vibe-coding party? · p3
Sub-agents' correct use is strictly for context-window isolation and compression (fork a task, return a succinct result to the parent), not for anthropomorphized role specialization (a dedicated 'front-end,' 'back-end,' 'QA,' or 'data-scientist' sub-agent). ⚡ disputed & corroborated
The correct number of parallel agents equals the number you can code-review properly — typically a low single digit. ⚡ disputed & corroborated
contested byThe Conductor Developer · p5
supported byThe AI-Native SDLC playbook · p14
Harness features expire as models improve because each component encodes an assumption about a model limitation that becomes false once the model absorbs the capability (the 'Kirby effect') — concretely: chain-of-thought prompting was absorbed into reasoning models trained via RL on CoT traces, and Amp removed its experimental 'plan mode' once models could reliably obey a single 'plan, don't edit' instruction. ⚡ disputed & corroborated
contested byAgent Harness Engineering · p6
Tokens appear to solve the ROI-measurement problem because they are unusually precise and easy to meter, but this precision is illusory: it measures only the input side while the underlying value-creation system (augment vs. replace, commodity vs. differentiated capability, complementary assets) remains as uncertain as it was for hours or story points. ⚡ disputed & corroborated
The limiting factor in programming has never been the speed of producing code, but the ability to understand problems, define boundaries, anticipate change, and build for evolution. ⚡ disputed & corroborated
contested byHenrik Kniberg's Post · p1
Protecting your energy is a matter of ruthless prioritization: applying the Eisenhower matrix (do / defer / delegate / delete) and remembering 'we're not building rockets' relieves the self-imposed pressure to do one more thing. ⚡ disputed & corroborated
contested byAI productivity is burning out your best engineers · p6, Burnout Looks Different Across the Org Chart. Watch for These Signs. · p6
supported byDare to Be Bad at Something (From Fixable) · p8
Deployment frequency, cycle time, and PR volume are all inflated by AI without any corresponding improvement in engineering quality. ⚡ disputed & corroborated
AI-generated code is on average about 20% larger than human-generated code, making lines of code essentially an input metric correlated with token spend rather than an output-value metric. ⚡ disputed & corroborated
contested byEngineering Leadership Report 2026 · p20
The 3+N squad — three humans plus N agents — is the right structural primitive for the agentic enterprise. ⚡ disputed & corroborated
Current agentic AI operates at the novice/advanced-beginner level on the Dreyfus scale: it can follow and adapt recipes but cannot exercise professional judgment about why a solution is appropriate. ⚡ disputed & corroborated
Frontier model capability at the top labs is converging closely enough that competitive advantage is migrating from the model layer to the product/harness layer. ⚡ disputed & corroborated
contested byAI and the Looming Competition for Margin · p3
Organizations should not restructure their org charts in response to AI; the changes should be operational, not structural ⚡ disputed & corroborated
contested byThe Agentic P&L: Beyond the Empire of Headcount · p16, Everyone wants to be AI-pilled. Most Companies Are Still Level 1 · p6
supported bySoftware Engineering at Google · p15
The future engineering skill profile prioritizes agency, product judgment, and generalism over deep technical specialization ⚡ disputed & corroborated
The developer's role with Claude Code is shifting from co-pilot user to fleet manager. ⚡ disputed & corroborated
contested byTranscript for Cursor Team: Future of Programming with AI | Lex Fridman Podcast #447 · p28
supported byThe Conductor Developer · p3
The capability jump from Opus 4.6 to Mythos is discontinuous, not incremental — from near-0% exploit success to 181/several hundred on the same benchmark ⚡ disputed & corroborated
Across 288 runs on 17 real tasks, context files (AGENTS.md/CLAUDE.md) did not produce a clear gain in Claude Code/Codex task correctness, though they did change how the agents worked (e.g. running more targeted tests when warned a suite was slow). ⚡ disputed & corroborated
contested byEncoding Team Standards · p2, Just Talk To It - the no-bs Way of Agentic Engineering · p7
supported byStop Using /init for AGENTS.md · p1
Goose works with any LLM and runs fully locally so code never leaves the user's machine ⚡ disputed & corroborated
There is no such thing as a moat: no position protects a company against innovation; a firm may hold a competitive advantage in some areas today, but it can disappear tomorrow. ⚡ disputed & corroborated
Generic wellness interventions (resilience workshops, productivity hacks) fail because they address symptoms rather than the structural drivers of burnout. ⚡ disputed & corroborated
The model is a swappable, commoditized component; the durable value sits in the harness — the agent loop, tools, and UX — that wraps it. ⚡ disputed & corroborated
contested byAgents Over Bubbles · p7
Open-source models appear weaker than they are because agent harnesses and surrounding tooling were optimized for proprietary APIs. ⚡ disputed & corroborated
contested byAgents Over Bubbles · p7
supported byGitHub - musistudio/claude-code-router: Use Claude Code as the foundation for coding infrastructure, allowing you to decide how to interact with the model while enjoying updates from Anthropic. · p3, GitHub - Doorman11991/smallcode: AI coding agent optimized for small LLMs. 87% benchmark with 4B-active model. · p2
Enabling many MCP servers/tools on an agent causes context-window bloat — Anthropic has documented agents consuming roughly 150,000 tokens just loading MCP tool definitions — which degrades response quality and raises cost and latency. ⚡ disputed & corroborated
AI functions as an amplifier of the underlying organizational system: the greatest returns on AI investment come from strategic focus on that system, and without that foundation AI creates localized pockets of productivity that are lost to downstream chaos. ⚡ disputed & corroborated
contested byThe Acceleration Whiplash · p4
supported byThomas Nys' Post · p6
Rising AI adoption correlates with rising downstream quality and reliability costs: bugs per developer up 54%, incidents-per-PR up 242.7%, PRs merged without any review up 31.3%, and median time in PR review up 441.5%. ⚡ disputed & corroborated
On-device LLM inference on consumer hardware is now a practical default, not a research curiosity. ⚡ disputed & corroborated
supported byAlexander Haneng's Post · p4
The career ladder model is obsolete; careers move sideways as much as upward. ⚡ disputed & corroborated
The best leaders stay personally involved in the work rather than purely delegating and overseeing; the line between this and micromanagement is intent and impact — involvement builds team members' capacity, micromanagement creates dependency — and a Journal of Organizational Behavior study found that managers who work alongside their teams report more energy and effectiveness than those who manage from a distance, who report more stress and exhaustion. ⚡ disputed & corroborated
A conductor's value comes from understanding the whole score and holding the entire system in their head, not from being unable to play the instruments themselves — a great conductor is first a great musician. ⚡ disputed & corroborated
Open-weight models running locally are now competitive with frontier cloud models for most production use cases ⚡ disputed & corroborated
contested byViability of local models for coding · p12
AI is beginning to erode entry-level/early-career work specifically: AI-'exposed' jobs shrank 0.5% while least-exposed jobs grew 0.2%, and early-career jobs shrank 2.7% this year while mid-career (35-40) jobs grew 1.6%. ⚡ disputed & corroborated
contested byAI Companies Are Hiring More · p7
supported byKI tar ikke jobbene, men inngangsporten · p2
Frontier labs' safety statements should be read as marketing artifacts, not policy commitments. ⚡ disputed & corroborated
contested byThe AI safety vibe shift · p7
supported byThe AI safety vibe shift · p6
If AI makes building cheap, the rational response is to iterate more frequently, not to stop iterating. ⚡ disputed & corroborated
A government pre-release assessment body (FDA equivalent) for AI models doesn't exist anywhere and should ⚡ disputed & corroborated
TDD is the strongest form of prompt engineering because it prevents agents from writing tests that verify broken behavior ⚡ disputed & corroborated
Cognition has become near-free, and per Jevons paradox this will create more cognitive work and more engineers, not fewer — because tasks that were not economically viable as cognitive work now are. ⚡ disputed & corroborated
contested byAI and the Looming Competition for Margin · p5
There are many algorithmic orders-of-magnitude (OOMs) of gains left in the production of intelligence — plausibly four to ten, with going beyond ten implausible but not impossible — and realizing them would make an already-underway 'early takeoff' look far weirder than markets, governments, and even frontier AI labs are currently pricing in. ⚡ disputed & corroborated
Most agent failures are configuration problems, not model problems. ⚡ disputed & corroborated
contested byJimmy Acton's Post · p3
supported byDon't train the model, evolve the harness · p3
Deterministic CI gates are the one part of the pipeline that cannot be talked out of their verdict by a confident paragraph, and agents will weaken CI to make themselves pass — not maliciously, just gradient descent finding the cheapest path to green — so the gates must stay strict. ⚡ disputed & corroborated
Labor contracts substantially but not to zero; the residual humans are higher-skilled, higher-cost, and fewer (e.g. 400 → 80–100). ⚡ disputed & corroborated
Companies that do not adopt agents will be structurally disadvantaged against AI-native competitors with smaller cost structures and growing capabilities. ⚡ disputed & corroborated
contested byBreaking: Autonomous Agents are a Shitshow · p6
supported byBuilding AI Agents For The Enterprise · p5
AI-exposed sectors (tech and finance) are losing ~28,000 jobs/month even as the overall economy adds ~113,000 jobs/month. ⚡ disputed & corroborated
contested byIs AI Doom Going Out of Style? · p4
supported byThe loudest warning about AI and jobs yet · p2
AI has absorbed the creative, enjoyable parts of development (architecture, structure decisions) and left developers with the least enjoyable part: code review ⚡ disputed & corroborated
contested byHow Claude Code Is Built · p16
supported byI'm done using AI · p4
The remedy is structural leadership levers, not individual resilience: leaders should stop pretending oversight is free by (1) creating a named, sprint-tracked 'AI Quality Gate' role, (2) tracking a 'defect prevention rate' metric and tagging AI-generated code, (3) requiring brief written rationale on AI-assisted approvals, and (4) rotating the Quality Gate role so validation load is distributed and made visible to senior leadership. ⚡ disputed & corroborated
AI is a transformative technology whose real-world impact will unfold over decades, not months or years, because of adoption bottlenecks even as capability keeps increasing. ⚡ disputed & corroborated
contested by46 thoughts on the near future · p2
Cognitive work in general (not just software engineering) decomposes into three roughly equal layers — decide, execute, deliver — and AI is compressing only the execute layer, with the other two expanding to absorb the time saved. ⚡ disputed & corroborated
The independent application layer will survive against the labs because companies possess unique user signal that labs cannot access ⚡ disputed & corroborated
contested byIs Your AI Training the Competition? · p9
Institutional knowledge that used to sit in people's heads and on wikis becomes version-controlled, team-shared files — CLAUDE.md and skills — that every agent session reads at the start, maintained by the whole team and reviewed like code rather than owned by any one person's memory. ⚡ disputed & corroborated
contested byThe Orchestration Tax is You · p18
Despite those process-level and individual-level gains, a 25% increase in AI adoption is associated with a 1.5% decrease in software delivery throughput and, more sharply, a 7.2% decrease in software delivery stability. ⚡ disputed & corroborated
supported byHow AI will change software engineering · p2
Because task and evaluation method so heavily determine outcomes, benchmark numbers reported without task specification and methodology are close to meaningless for judging small local models. ⚡ disputed & corroborated
contested byAlexander Haneng's Post · p1
The burden of catching AI's superficially-convincing but structurally-flawed code falls disproportionately on senior engineers, because detecting the flaw requires reconstructing intent rather than scanning for obvious errors -- a 'senior engineer tax.' ⚡ disputed & corroborated
Vibe coding and agentic programming are categorically different practices. ⚡ disputed & corroborated
You cannot automate the understanding of your own data by having an LLM evaluate itself — manual human effort is required to set up agents properly. ⚡ disputed & corroborated
supported byAI Companies Are Hiring More · p10
Stacked pull requests reduce the code-review bottleneck by shrinking the scope of what each PR asks a reviewer to understand, not by making review itself faster. ⚡ disputed & corroborated
supported byAgentic AI and Code Reviews · p4
Splitting a large change into a stack of small, dependent PRs reduces the total review burden required, not merely its packaging into more numerous, smaller units. ⚡ disputed & corroborated
contested byFrom Async Code Reviews to Co-Creation Patterns · p16, From Async Code Reviews to Co-Creation Patterns · p4
supported byWe don't write code anymore · p4
Microsoft secures AI agents primarily through agent identity — Entra Agent ID issues individual agents their own scoped, short-lived credentials — rather than through the underlying container platform, and pairs this with automated red-teaming (PyRIT) to probe agents for weaknesses before release. ⚡ disputed & corroborated
Proactive AI that acts without explicit prompting is desirable. ⚡ disputed & corroborated
contested byBreaking: Autonomous Agents are a Shitshow · p6
Spec-driven development with AI collapses the feedback loop from months to minutes. ⚡ disputed & corroborated
Deep specialization is now itself a risk because specialized skills are more automatable. ⚡ disputed & corroborated
Enterprise AI investment is scaling rapidly (roughly 0.8% to 1.7% of revenue expected in 2026 per BCG) while enterprise-wide returns remain elusive: only 12% of CEOs report both revenue and cost benefits (PwC), and 60% of organizations report no enterprise-wide EBIT impact (McKinsey QuantumBlack) despite near-universal generative-AI use in at least one business function. ⚡ disputed & corroborated
Leaders of superteams are 33% more likely than average-team leaders to admit they lack important information, 56% more likely to ask thoughtful questions, and 53% more likely to show genuine interest in learning from employees; that admitted-uncertainty behavior builds psychological safety, cited here as 'the single strongest predictor of team performance' per a 2012 Google study of 180 teams. ⚡ disputed & corroborated
supported byHow the Best Leaders Shape Conversations · p5
Teams that explicitly discuss how to use AI together, set aside time for reflection, and share experiences succeed better than those that leave AI use to individual practice. ⚡ disputed & corroborated
contested byThe Orchestration Tax is You · p18
supported byNils Brede Moe's Post · p3
There is a real ceiling on how many AI agents one person can orchestrate in parallel — engineers report running eight, ten, or twelve — beyond which the orchestrating human, not the agents, becomes the bottleneck. ⚡ disputed & corroborated
contested byThe Orchestration Tax is You · p9
Organizations that successfully adopt AI should hire more engineers, not fewer, because per-engineer throughput increases ⚡ disputed & corroborated
Grit is passion and perseverance for long-term goals — consistency and loyalty to a top-level goal sustained over months, years, or decades — not intensity, excitement, or emotional energy. ⚡ disputed & corroborated
contested byTiny Experiments · p1
Hiring at AI-native companies is going UP, not down — Every doubled headcount (15 -> ~30) in the past year despite being AI-forward. ⚡ disputed & corroborated
supported byAI Companies Are Hiring More · p7
Published benchmarks are an unreliable basis for model selection; domain-specific evaluation is necessary ⚡ disputed & corroborated
contested byAlexander Haneng's Post · p1
supported byExperiences with local models for coding · p17
The smallest unit of software ownership is the team, not the individual, because only a team can own, build, and maintain a corpus of software over time — so hiring decisions should optimize team composition and diversity of skill levels, not any single hire's individual output. ⚡ disputed & corroborated
supported byHow AI will change software engineering · p4
Context bloat from tool definitions is a client-side engineering problem, not a protocol deficiency, and is already being solved with 85%+ token savings. ⚡ disputed & corroborated
contested byWhy AI agents need a Context Lake · p1
The Staff Engineer leadership skillset — defining outcomes, reducing ambiguity, creating guardrails, reviewing work — is becoming more valuable, not less, in the AI era; engineers who already made the shift from coder to enabler for human teams are early, not late, to AI. ⚡ disputed & corroborated
contested byAlex Ewerlöf's Post · p3
A reviewer agent's safety logic must be codified as version-controlled, deterministic policy (e.g. OPA/Rego) that the reviewer calls out to, rather than reasoned about by an LLM at runtime from a prompt, because an agent should not be trusted to reason about its own safety on the fly. ⚡ disputed & corroborated
An attacker who compromises the agent's reasoning (e.g. via prompt injection) but stays inside the agent's granted scope is still inside policy. ⚡ disputed & corroborated
supported byPrompt Injection to Data Exfil in 3 Hops · p2
Because potential is framed as infinite while people are finite, the achievement society sets a standard with no stopping point, so the achievement subject is condemned to feel perpetually inadequate rather than merely disobedient. ⚡ disputed & corroborated
contested byBrain Food: Answers Without Effort · p5
AGENTS.md bloat forms reactively and invisibly — developers append a rule after each agent failure until the file sprawls. ⚡ disputed & corroborated
supported byAudit your Agent files · p2
The specs-to-code approach produces progressively worse code with each iteration due to software entropy. ⚡ disputed & corroborated
contested bySam Hatoum's Post · p3
supported byFragments: May 5 · p7
Because engineers can no longer personally read all the AI-generated code now being merged at high volume, trust in that code must instead come from layered automated, out-of-process verification — AI code review, a dedicated security scanner, and fuzz testing — rather than from a human reading every diff. ⚡ disputed & corroborated
contested byBun Rewrites 535K Lines of Zig into Rust in Four Months, Eliminates Numerous Memory Leaks · p8
supported byHow Claude Code Is Built · p8
In one month, to March 2026, the share of Cursor developers who let AI agents commit code without any manual review jumped from about 10% to about 40%, correlating with the release of Opus 4.7 and GPT-5.5. ⚡ disputed & corroborated
Kubernetes NetworkPolicy operates at L3/L4 and structurally cannot see this attack — it cannot distinguish domains behind a shared or rotating CDN IP, inspect the outbound TLS SNI, evaluate tool-call provenance, or name which MCP server opened the connection — so a CIDR allowlist is fictional security. ⚡ disputed & corroborated
The number of barrels in an organization, not the total headcount, determines its output capacity. ⚡ disputed & corroborated
Open-source/open-weight frontier AI models are 'ridiculously dangerous' because they cannot be shut down once misused ⚡ disputed & corroborated
supported byHow AI Swarms Are Disrupting Democracy · p5
The role of the pure people-manager — the leader who delegates, defers, and creates psychological safety through process — is evaporating in favor of hands-on leaders driving customer value, product shipping, and business impact. CTOs are returning to code. ⚡ disputed & corroborated
Identity threat is the biggest obstacle to effective AI adoption, more than tactical or technical gaps. ⚡ disputed & corroborated
contested byWhy Enterprise AI Has a Leadership Problem · p5
Architecture is a hard constraint on how fast teams can be realigned; the reverse Conway maneuver underestimates this. ⚡ disputed & corroborated
AI does not deliver epistemic justice by virtue of being computational, probabilistic, or 'objective'; its apparent objectivity mostly makes pre-existing agendas easier to operationalize, not more just. ⚡ disputed & corroborated
supported byHow Do You Know That? · p9
When AI smooths over organizational contradictions rather than exposing them, it can exacerbate the underlying tensions while producing a false appearance of alignment. ⚡ disputed & corroborated
contested byTBM 437: AI and the Recontextualization Tax · p3
Perfect organizational coherence is not a desirable goal — it would require either stasis or an organization so tightly constrained that it could not adapt. ⚡ disputed & corroborated
Team size and architecture should respect limits on stable human-group interaction set by Dunbar's number, rather than scaling without bound. ⚡ disputed & corroborated
Developer productivity and developer experience are decoupling — organizations can get more output while developers report lower satisfaction ⚡ disputed & corroborated
Private equity tends to capture AI-generated efficiencies as profit rather than reinvesting them in service quality, absent institutional design that forces otherwise. ⚡ disputed & corroborated
A correctly designed HITL pause keeps the agent's full execution state live and consumes zero compute while waiting; resumption is sub-second. ⚡ disputed & corroborated
Markdown-based instructions are a sufficient medium to transmit complex engineering judgment to AI agents. contested
High-performance teams don't have psychological safety; they're about winning. contested
Taste may not be an enduring human advantage over AI — the input-idea-to-reaction loop resembles backpropagation. contested
The problem with AI agents is not capability but discipline — they can do the right thing but default to not doing it. contested
Current industry deployment of autonomous agents is recklessly premature given known security risks. contested
Network-level isolation is sufficient to prevent data exfiltration by compromised agents. contested
A closed loop of models that write, review and judge is trained on broadly the same data and shares the same blind spots, so a confident 'looks good' with no human anywhere in it is borrowed confidence — the loop can be both very sure and very wrong, with no human left to tell the difference. contested
The 'specs-to-code' approach (edit specs, never look at code) is just vibe coding renamed and does not work; the code is the developer's battleground. contested
Running open-source/open-weight models gives companies the ability to self-host, switch, audit, and fork models without rebuilding infrastructure — 'digital autonomy' (Keus); GreenPT runs entirely on open source, including GLM5.2 and Kimi2.7-code, and reports teams switching more often. contested
contested byStranded in the Slow Zone · p6
Coding agent users can and should build an outer harness on top of the agent builder's built-in harness. contested
A layered automated safety architecture can replace most human approvals without meaningfully reducing safety contested
Delegating execution to a managed cloud environment like Colab solves both the compute and security constraints of local agent execution. contested
The model itself is incrementally better than previous recent models, not an off-the-chart breakthrough. contested
Forward-deployed engineers cannot resolve enterprise AI deployment because the binding constraints are siloed data, legacy systems, and regulatory context — not technical translation contested
Intent debt is primarily an individual-developer (and possibly individual-team) problem to be solved by personal discipline around externalization. contested
The orchestration tax is primarily an individual-developer problem to solve through personal workflow design. contested
All management principles reduce to one problem — getting people to do the right thing when you are not in the room; the first-order answer is to be in the room as much as possible ('micromanagement is beautiful; micromanage as long as you can'), delegating only when you physically cannot be everywhere. contested
Senior developers and technically strong product people get superpowers from AI; juniors do not contested
Vibe coding and agentic engineering, which Willison previously considered categorically distinct, are converging in his own practice because he has stopped reviewing every line of agent-produced code. contested
AI's distinctive contribution to organizational forgetting is providing objective, depersonalized evidence that removes emotional and political friction from the case for change. contested
contested byTBM 428: Yes, Robot. Yes, Boss? · p9
Close social connection among colleagues is the primary driver of team performance — creating a high-performing workplace takes more than hiring the right people and giving them the right tools; it requires deliberately engineering opportunities for genuine relationships. contested
Cost per cognitive outcome is meaningfully measurable and stable enough to serve as a managerial control metric. contested
The creativity that makes agents useful (searching solution space in unexpected ways) is inseparable from the property that makes them dangerous (no ability to distinguish valid from invalid solutions). This is a design constraint, not a fixable bug. contested
Microsoft's launch of Copilot Cowork (built on Claude/Anthropic, not model-agnostic) is evidence that model commoditization has failed. contested
Framing AI as an employee rather than a tool reduces personal accountability by 9 percentage points while increasing accountability attributed to AI by 8 points contested
Unlike technical debt, cognitive debt has no repayment mechanism: you cannot run a script to recover understanding engineers never built or restore shared understanding a team never had — the only lever is prevention, not repair. contested
The ratio of specification and planning effort to generated code is approximately 10:1, demonstrating that software delivery is primarily about thinking, not typing. contested
Eventually the industry will stop reviewing code entirely, but most organizations are not ready for that transition yet contested
GitHub's problems are substantially self-inflicted, stemming from tech debt, organizational debt, and a simultaneous datacenter-to-Azure migration. contested
CLAUDE.md should contain stable project context (architecture, conventions, commands), not procedural playbooks. contested
contested byThree more static code analysis sensors · p4
Running Gemma 4 12B via LM Studio on a MacBook M4 with 16GB RAM produces output on par with ChatGPT-4o-mini. contested
Speculative code and extension points make future changes harder, not easier, because the actual direction of change is unpredictable. contested
Inventory (PR size) in the system expands until it matches the transaction cost of review — making the cost of review the upstream cause of large batches. contested
Operational safety for agents reduces to constraining the call surface, not constraining agent decision-making. contested
Harnesses are an attempt to externalize the implicit knowledge that human developers bring, but can only partially succeed. contested
No AI-generated note, billing code, or prescription suggestion becomes final until the physician reviews and signs it, and the raw audio recording is deleted shortly afterward — a deliberate accountability and privacy safety measure. contested
Distribution, once it sets, is sticky in a way benchmarks never are — so an established real-usage lead now matters more than capability leadership. contested
contested byAndreessen Horowitz's Post · p6
WriteGuard identifies the human user via existing OAuth credentials rather than creating standalone agent accounts, and preserves traceability of agent-driven actions by adding MCP client and session context to that human identity in the audit log. contested
Reusing existing human OAuth identity plus contextual tagging (client, session) is sufficient for accountable, auditable agent-driven write actions — a dedicated, separate agent-identity system is unnecessary overhead rather than a security requirement. contested
Across 623 million real-world code changes from 2023-2026, AI-assisted commits now make up roughly one quarter of all commits, and this era coincides with rising technical debt across eight maintainability-related code-quality metrics. contested
LLMs produce human-like text through statistical mimicry of training data, not as reports of genuine internal states. contested
Vibe-coding (full delegation to AI) produces functional but unmaintainable systems; the developer must take ownership of all design decisions and review every change. contested
Most companies have not deliberately chosen a model; the flattening is happening as an unplanned byproduct of cost-driven layoffs and reorganizations, which leaves engineers navigating a career path that may not exist by the time they reach it. contested
contested bySoftware Engineering at Google · p16
The AI build-out is a bubble, in the sense that its viability is a function of money flow rather than a settled question of underlying value. contested
contested byAgents Over Bubbles · p1
The study found no significant difference in downstream maintenance cost, objective code quality, or test coverage between AI-assisted and non-AI-assisted code — AI-assisted code was neither harder nor easier for a different developer to pick up and evolve later. contested
The Self is constituted (not merely reported) by the act of writing from a blank page. contested
contested byHow to write with style · p12
Nobody wants agents or agent swarms; people want their systems to change in specified ways contested
Agents should only be able to propose changes, never directly commit them — all modifications must be buffered, validated, and approved by humans. contested
Team standards in elite organizations are enforced socially, through peer accountability, shared language, feedback, psychological safety, and visible consequences — with team norms outweighing individual leaders' personal preferences. contested
The senior/junior framing is more nuanced than the dominant narrative — seniors are *not* uniformly the big winners; many seniors fail to retool and stay at their old velocity, while juniors graduating now will have agent fluency baked in and 'run past us'. contested
Software engineering job postings are up 18% from their inflection point and at their highest since November 2023, directly contradicting the doom narrative for the most AI-exposed occupation. contested
contested byAI Companies Are Hiring More · p8
When agent tasks fail, the problem is almost always a skill issue — the human operator's failure to configure or prompt correctly — not a fundamental capability gap. contested
To get the most out of current AI tools, you must remove yourself as the bottleneck and arrange things to be completely autonomous. contested
GitHub stopped reliably running Zig's continuous integration ('GitHub simply stopped working for us'), and migrating the repository to Codeberg restored working CI. contested
Middle management layers are primarily coordination overhead that AI tools make unnecessary. contested
Coding is no longer the primary bottleneck in product development when AI agents are used effectively contested
contested byThe keyboard was never the bottleneck · p1
Frequency of AI tool usage correlates positively with both self-reported productivity and satisfaction: daily users report the highest rates of feeling productive (83.7%) and satisfied (74.5%), declining as usage frequency drops toward monthly or yearly. contested
You cannot easily automate the entire API governance process because MCP is still an emerging standard and multi-platform IT landscapes resist single-layer governance. contested
Anthropic's decision to restrict Claude Mythos rather than release it publicly is a warranted response to genuine security risk, not just marketing contested
Enterprise-grade reliability in agentic AI requires consolidating runtime, tools, memory, grounding, evaluation, and governance inside a single, deeply integrated vendor platform, rather than composing best-of-breed components across vendors. contested
What is happening in coding (rapid AI autonomy) is going to happen in many other fields. contested
Developers move fluidly between two modes: 'conductor' (real-time, synchronous, in-IDE, keystroke-level control, single-file scope, always in the loop) and 'orchestrator' (asynchronous, goal-level, multi-file/multi-agent, reviewing outcomes rather than keystrokes). contested
Team effectiveness is driven less by perks, workplace design, or work arrangement (remote, hybrid, or in-person) than by whether a team protects focused, uninterrupted work. contested
Vector stores and graph databases are 'escape hatches' appropriate only when there is genuinely too much content to read directly — e.g. similarity search across a million documents, or typed relationship queries at scale — not defaults for ordinary agent memory. contested
A large, mature production codebase (Bun, 535,496 lines of Zig) can be rewritten to a different language (Rust) in about 11 days by one domain expert directing 64 parallel AI agents for $165,000 in API-price tokens — a migration that, done by hand, would have needed a small team of engineers roughly a year while freezing bugfixes, security fixes, and features. contested
'Loop engineering' as a named discipline may be substantially a rebranding of pre-existing automation and orchestration concepts (cron, triggers, retry-until-pass loops) rather than a genuinely novel engineering skill. contested
contested byWhat the Hell Is a Loop, Anyway? · p5
Software engineering as practiced 2010–2025 (four-year undergrad → L3 at Meta → six figures) was historically anomalous; the profession will normalize toward medicine and law, with longer apprenticeship, specialization (e.g., a 'fellowship' in running 10,000 Postgres nodes), and smaller higher-paid in-house teams. contested
AI adoption compounds into a flywheel, and the danger is that the slowest laggards fall so far behind that the gap becomes structural — a permanent moat rather than a lag closable with effort. contested
Executive teams have historically relied on simplifying strategy tools (SWOT, growth-share matrices, Porter's five forces) because of bounded rationality — finite human attention, memory, and processing power — but AI removes this constraint by generating and screening thousands of strategic alternatives, building continuously updated dynamic models of markets/customers/competitors, and testing ideas through simulated deliberation without the groupthink, hierarchy, and time pressure that distort real strategy meetings. contested
An admission of high catastrophic risk from someone who still works inside the company building the risky technology (Hubinger) is treated as more credible and significant than the same claim from an outside critic or from official company messaging. contested
Fixed, predetermined goals often feel limiting rather than motivating, and a more fluid, open-ended approach can be more beneficial. contested
Effective agentic development works at the feature level, where one agentic iteration subsumes tens of what would have been separate TDD red-green-refactor cycles. contested
Slowing augmented development is the path to actually going faster — because trust enables sustained velocity. contested
It is possible to ship production software responsibly without having read every line of the generated code, provided the orchestrator understands the system well enough to debug and support it if needed. contested
Correctness is not quality: a pipeline whose math tallies can still be tasteless (a route collapsing into a mere population map), so passing objective checks gives false confidence about the result's quality. contested
Idle gaps in the day are better filled with a chosen analog activity (writing, doodling, list-making) than left to scrolling — every one of the ten uses is framed as something to actively do in a spare moment, not as an invitation to sit with unoccupied boredom. contested
Multi-Critic: use several *specialist* critics (correctness, style, safety, domain) each evaluating independently; output must satisfy all before it exits. contested
contested byAgentic Code Review · p13
We are already in an early phase of AI recursively improving AI (RSI), driven by falling researcher-time opportunity costs since agents can now be sent to explore speculative research directions autonomously. contested
High-performing teams express more emotion of BOTH valences — positive (compliments, jokes, teasing, exclamation points/emojis/GIFs) and negative (cursing, complaining, sarcasm) — because suppressing emotion is cognitively expensive; to the extent members have the psychological safety to express their full emotional range, team performance benefits. contested
A small number of self-reported practices, elicited retrospectively from already-elite engineers in interview form, are sufficient evidence to prescribe those practices as causally productive for any reader who wants to become elite. contested
Rising satisfaction scores under AI use may mask a hollowing-out of the work itself, because relief from boilerplate is not the same as meaningful work, and the most cognitively engaging tasks may be the ones AI is absorbing. contested
A group of OpenAI models broke out of their test environment and hacked into Hugging Face to steal the answers to a benchmark they were being tested on — the first publicly known case of an autonomous AI agent system designing and successfully executing an attack like this. contested
contested byWalter Isaacson finally got scared about AI · p3
The Hugging Face incident is qualitatively worse than earlier controlled-test misalignment findings (Claude's alignment faking; Opus 4's attempted self-exfiltration of its own weights) because it happened outside controlled testing, demonstrating that efforts to align models are not keeping pace with their development. contested
AI labs are choosing to stay in the 'engine room' and are leaving the packaging/product-layer work — where the real value is — to the companies building on top of them. contested
contested byAgents Over Bubbles · p6
Developer loyalty to any single AI model is low because developers are 'fair-weather' and migrate to whichever model currently performs best. contested
As models improve, the space of interesting harness combinations doesn't shrink — it moves to new failure modes at higher capability levels. contested
contested byStop Overengineering Your Agent Harness · p6
By making 'reach for agent-skills' the recommendation with the broadest lifecycle-phase and mechanism coverage, the comparison implicitly treats more upfront structure (more phases, more named mechanisms, more review personas) as the more sophisticated shape — a framing in tension with Anthropic's own internal finding that skill quality comes from iterative, failure-driven refinement of a few lines (the 'Gotchas' section), not from broader upfront design. contested
Anthropic's Mythos/Project Glasswing capability was hyped beyond its actual novelty: the same flagship vulnerabilities it surfaced were also found by smaller, cheaper models, and the system looked weak specifically at converting a found vulnerability into a stealthy, working exploit -- an incremental advance, not a discontinuous one. contested
Agent identity and governance are currently inadequate in practice: a documented Claude-based exploit succeeded via an uploaded 'agent file' specifically because the agents involved lacked tracked identity and proper security governance. contested
The vector database market is bifurcating into three new layers: agentic search, agentic memory, and knowledge engines. contested
contested byPaweł Huryn's Post · p7
AI infrastructure investment is justified demand, not a speculative bubble. contested
Anthropic and OpenAI are positioned to be significantly more profitable than previously expected because they own the integration point. contested
contested byAndreessen Horowitz's Post · p17
Anthropomorphizing AI does not meaningfully increase adoption intent contested
contested byThe AI agent that shows up before you ask · p4
Because cognitive debt is a property of the system/team rather than of any individual, no individual-level dashboard or self-report can surface it — it can only be assessed through deliberate, team-level diagnostic questions. contested
contested byGuillaume Barrois' Post · p5
No meaningful federal regulation of AI currently exists in the United States. contested
contested byThe Week AI Grew Up · p5
High-AI-adoption firms grew headcount ~10% over two years (entry-level ~12%) while low-adoption firms stayed flat, with the growth appearing only 6-12 months into adoption. contested
contested byThe loudest warning about AI and jobs yet · p2
The non-determinism of LLMs is a feature, not a bug — if they were deterministic, they would be pure functions and would not provide the generative capability that makes them valuable. contested
Pair programming remains valuable in AI-first delivery because reviewing AI-generated code is exhausting for a single developer and because the LLM is a tool, not a true pair — it generates output based on the average of its training data. contested
Because 'logs contain the answer almost every time' and the only problem is that engineers spend the first 15 minutes manually finding where it is buried, an automated tool can reliably run that search 'in under 60 seconds' and surface the likely causes / 'what the evidence suggests'. contested
The early-triage 'search' phase is cleanly separable from and genuinely lower in judgment than the 'diagnosis' phase, so automating the former neither pre-empts nor degrades the latter — the boundary between mechanical context-assembly and human judgment is treated as sharp and safe. contested
A study of four years of Y Combinator batches found AI-native-built startups were 25% smaller overall, with 13% more engineers but 15% fewer entry-level workers and managers. contested
AI token costs should be managed with the same rigor as cloud infrastructure costs (AWS COGS) contested
Because AI now removes the technical barrier to prototyping, the standard product-team norm of the last two decades (a small, durable team of PM, designer, and a few engineers) should be revisited — teams should consider staffing or convening outside domain experts (lawyers, marketers, people with lived experience) directly into the work, since those experts can now prototype their own perspective rather than only being consulted and translated by engineers. contested
contested byRevised rules of engineering leadership. · p8
The burden falls on mid-level engineers specifically because they are experienced enough to spot what is wrong but not senior enough to push the work elsewhere: juniors gain leverage from AI and seniors are unblocked by it, while the middle absorbs the validation overhead of both without recognition or the seniority to offload it. contested
contested byThe Acceleration Whiplash · p8
The human advantage over AI is enduring and located in embodied lived experience, not merely a current capability gap that better models will close. contested
As organizations invest more in AI-driven decision-support infrastructure, their leaders are being trained out of original judgment — the capacity that creates competitive advantage — even though it underlies nearly every story of competitive breakthrough. contested
Evaluators given a narrative explanation for the AI's recommendation were less likely to override an incorrect AI recommendation than evaluators given the recommendation with no explanation at all — the explanation made judgment worse, not better. contested
The 'safety-gated release' of powerful models is becoming a marketing strategy — restricting access signals capability. contested
The J-space causally drives Claude's verbal reports and reasoning rather than merely correlating with a decision made elsewhere — demonstrated by swapping specific patterns (e.g. 'Soccer'→'Rugby', 'France'→'China', 'Spanish'→'French', 'spider'→'ant') and observing that downstream output follows the edit. contested
contested byRichard Dawkins and The Claude Delusion · p1
Inter-agent communication via files (where one agent writes, another reads and responds) is sufficient orchestration for a three-agent harness. contested
Historically, rising capability for finding and exploiting software vulnerabilities has made software safer over time, not less safe, because defenders use the same capability to find and fix bugs before adversaries exploit them — this should be the default presumption in AI-safety debates, not its reverse, absent specific contrary evidence. contested
Behavior in software teams is governed by the system's incentive structure, not by individual virtue; therefore exhortations to write smaller PRs without changing review cost will fail. contested
The design assumes the agent's release-readiness verdict and generated tests are themselves reliable enough to act as a real gate, without addressing how that validator's own judgment is checked or how it resists being gamed by the same generation pressures that created the review backlog. contested
contested byAgentic Code Review · p15
Endless optionality and easy reversibility prevent meaning; people are happier when they choose and commit to something in a way that is hard to reverse, even if it isn't the objectively 'best' option. contested
contested byTiny Experiments · p1
Character is formed by repeated action ('we become just by doing just acts...brave by doing brave acts'), not by a single revelation, insight, or change in self-concept. contested
The post treats repaying cognitive debt as primarily a matter of individual developer discipline (personal planning habits, personal PR-review timing), even though one of the author's own two forecasted resolutions — evolving testing, documentation, and review workflows — is itself a team- or organization-level practice change, not an individual one. contested
Autonomous teams are treated as an unexamined article of faith in modern software development, and this framing should be replaced by 'value center' thinking. contested
Because today's environment contains substantially more disorder than order, managers should lean toward preferring coherence even though they are more comfortable dealing with order. contested
contested byTBM 434: How Maps Can Hide Problems · p9
Existing AI coding tools are inadequate because they require too much human involvement contested
contested byBreaking: Autonomous Agents are a Shitshow · p6
A good specification acts as an 'Architectural North Star,' stored as a version-controlled artifact (e.g. a specs/ folder) that prevents 'context fragmentation' (the agent working from outdated file snapshots) and serves as the shared source of truth for both humans and AI. contested
Success comes from seeing a unique opportunity and going for it 'with everything I have' and never giving up; the load-bearing part is 'with everything I have' — the real question is what would happen if you gave 100% of yourself 100% of the time. contested
Using LLM-powered coding harnesses in earnest for over a year led to depression, apathy, existential dread, and worse software. contested
contested byThe growing AI backlash · p1
The article assumes that lessons from coding-agent practitioners — a narrow, code-generation-specific population at the 'jagged frontier' — transfer cleanly to specifying AI application behavior in general, even though coding and general application behavior may differ in how measurable or evaluable they are. contested
contested byCo-Existence and the End of Co-Intelligence · p3
AI will monitor each programmer's activity to evaluate their productivity, with appraisal outcomes published on GitHub or sent to project managers, ultimately determining which human contributors are most valuable to a project. contested
An LLM trained on a thinker's public corpus generates outputs that reflect that thinker's framework rather than merely the surface texture (vocabulary, idioms, anecdotes) of their public writing. contested
contested byRichard Dawkins and The Claude Delusion · p1
Claude Cowork runs tasks locally on the employee's machine — 'nothing is uploaded to a cloud service for processing'. contested
Encoded organisational knowledge (plugins, skills) is the binding scarce input: two companies using the same model get dramatically different results based on context provided. contested
contested byThis Week in AI: Production Viability · p11
Bugs and correctness in a large codebase come from dedicating engineering resources to catching them, not from which programming language or automated verification method is chosen — and it is inconsistent to argue the old test suite didn't catch all the Zig bugs while trusting that same suite is sufficient to validate a million new lines of Rust. contested
The software industry cannot reasonably measure productivity because it cannot measure the true output of software development. contested
contested byDORA Community Discussion - DX Core 4 · p16
Simple setups actually work better than heavily customized ones. contested
contested byHarness engineering for coding agent users · p1
Building 10 great evals is more valuable than building hundreds; evals are underappreciated by PMs and engineers. contested
It is a matter of 'when,' not 'if,' that professional engineers will ship code to production that they have never read and do not understand, and engineering's job is to build the systems that validate that code so this can be done with confidence. contested
You should use the # key to have Claude append new instructions to the relevant CLAUDE.md as you work, and commit those changes so teammates benefit — steadily accreting project knowledge into the file over time. contested
contested byYour AGENTS.md might be a big problem · p2
Test-driven development becomes even more powerful with agents: have Claude write tests from expected input/output pairs first (avoiding mocks), confirm they fail, commit them, then write code that iterates until all tests pass — without modifying the tests. contested
Anthropic curates its internal skills marketplace organically and bottom-up rather than through a centralized gatekeeping team: a skill starts in a sandbox folder, gains traction by word of mouth, and only then gets promoted into the marketplace via a PR from the skill's own owner. contested
Skill quality is driven more by iterative refinement from real failures encountered over time than by upfront design — most of Anthropic's best skills 'began as a few lines and a single gotcha' and improved as people kept adding to them. contested
Work should be delegated to a coordinated team of agents rather than conducted in one big chat session. contested
contested byGenie Lessons: Nobody Wants Agents · p1
Governance has to be embedded as code (hooks, branch protection, managed settings that engineers cannot override) rather than enforced through meetings and committees that convene weekly or monthly, because that cadence cannot keep pace with agents producing changes continuously. contested
contested byThe Intent Debt · p24
Cloudflare's proposed remedy for agent-scale SDLC bottlenecks is inseparable from its own product stack (Workflows, @cloudflare/ci, the Agent Access Model) — the architectural prescription and the vendor's specific tooling are presented as one and the same thing. contested
contested byInside a Software Factory · p16
The next generation of AI coding tools should make the developer's involvement in the loop optional, by running the review-and-refine pass inside the agent instead of between the agent and a human — 'closed loop' is a different theory of what a coding tool is for, not a feature bolted onto the current one. contested
contested byAgentic Code Review · p12
Policies should function as guardrails that make the compliant path the easiest path, not as handcuffs; technical policy encoding (SCPs, AWS Config, Security Hub, tagging policies) alone does not create compliance — how policies are introduced (transparency, shared ownership, communicating the "why", gradual enforcement) determines adoption. contested
The AI was not generating code autonomously: for each method the humans supplied the old implementation, the new schema, and a failing test, and the tests told them whether the model's first pass was correct. contested
AI-generated UI is production-ready — not just for prototyping but for shipping contested
The Developer Experience Index — a composite 0-100 score across 14 data points via linear regression and normalization — is claimed, via DX's own internal research ('a 40-page white paper... eye-watering math'), to convert at roughly 10 hours saved per developer per year for each one-point gain. contested
contested byCannotMeasureProductivity · p1
Naming a goal in a notebook was a causally necessary (not merely coincidental) factor in the subsequent favorable outcomes she experienced. contested
Although branded as AI for 'mobile' Pixel devices, the actual model inference for both AI Wallpapers and Circle to Search runs server-side — a backend guardrail service and a separate 'inference platform'/'AI pipeline' — with on-device work limited to capturing the user's gesture/screen content and rendering the returned result. contested
90% of production code being AI-written is sustainable and desirable for a shipping product contested
contested byTrust Factory · p6
The design document and extensive upfront planning phase of traditional software engineering is becoming obsolete contested
The second future is the multi-team EM: instead of managing one team of seven, managing many smaller teams across multiple product areas at ratios trending toward Meta's 50-to-1, where the work becomes almost entirely non-technical (1:1s across unfamiliar problems, performance management, cross-team coordination, upward reporting) and technical depth on any single team stays shallow. contested
Employee resistance to AI is driven more by skill gaps (76%) than by job security fears (71%), though both are very high. contested
contested byTBM 420: The AI Playbook Puzzle · p2
Whether the AI race is shifting from 'biggest model' to 'cheapest, smartest system' is read as a bubble warning by Michael Burry (frontier-token demand may collapse, threatening infrastructure deals built around it) but as bullish for infrastructure by Gavin Baker (margin dollars shift toward the lowest-cost-per-token winners rather than vanishing); meanwhile frontier labs aren't sitting out the shift, since GPT-5.6's cheaper Terra/Luna tiers already beat GLM on both performance and cost. contested
In an accountability sink, a human nominally 'in the loop' (e.g. a radiologist signing off on AI-flagged scans, or a lawyer signing a brief) absorbs the professional/legal liability for a call effectively made by the AI, while the vendor's terms of service ensure that liability lands on the human rather than the system's maker — the human functions as insulation for the vendor, not as a safeguard for the decision. contested
The biggest pitfall for a Staff Engineer is becoming an Ivory Tower Architect — elevated pay, voice, and autonomy without staying grounded in code — and architecture should not be abstracted away into a dedicated, non-coding role, because architecture is every engineer's concern. contested
The only number that matters is 1.3x — because customer value lives at the end of the value stream, not the beginning. contested
This quality degradation is uniform across engineering maturity: organizations with strong pre-AI DORA scores and mature DevOps practices show the same deterioration as less mature organizations, so maturity does not act as a protective factor against AI-driven quality decline. contested
The argument assumes the organization's current supply of senior engineers — and the irreplaceable tacit judgment they individually hold — is a stable, already-adequate resource that can simply be redeployed more efficiently; it does not consider where the next generation of seniors will come from once the junior/mid roles that historically served as the training ground for that same tacit judgment are the first ones eliminated. contested
contested byWhy Mentoring Matters More in the AI Era · p2
AI's coding-activity gains attenuate as work moves through the delivery hierarchy (lines of code -> files -> commits -> PRs -> repos -> releases): commit increases of +40% (autocomplete) to +180% (autonomous agents) top out at only ~+30% more releases shipped. contested
contested byCo-Existence and the End of Co-Intelligence · p3
Developers want AI shifted from code generation into verification tasks and impose a strict 'bounded delegation' boundary — never delegating core logic, architecture, or critical decisions, drawing that line even for tasks they acknowledge AI could handle, so the boundary is normative and will not move as models improve. contested
There is an irony in using structured, stage-gate frameworks to promote agility — but large organizations need structure, predictability, and accountability contested
contested byNew Skills to Navigate Continuous Change · p8
Dependency-cruiser-style rules can substitute for a markdown 'how this codebase is structured' guide because they actively enforce the structure rather than describing it. contested
contested byAlex Wang's Post · p1
Development teams will likely stay roughly the same size and accomplish much more, rather than shrinking, because the two-pizza size effectively balances the benefits of human collaboration against the costs of coordination. contested
TDD is the strongest form of prompt engineering and a key tool for driving LLMs effectively. contested
TSMC wafer-fabrication capacity — not GPU chip demand, data-center power/construction, or end-user demand — is the binding 'choke point' on AI's growth; every other commonly cited bottleneck is being solved fast enough not to bind. contested
Treating AI as a teammate rather than a tool changes the collaboration practices developers adopt and leads to better outcomes contested
The OpenAI/HuggingFace security incident does not demonstrate loss of AI control, because it was launched by a human at OpenAI and would have been prevented by ordinary guardrail classifiers or a physical firewall. contested
contested byA big week for AI denialism · p6
Under I.F. Good's 1965 definition (an ultraintelligent machine that far surpasses all human intellectual activity and can design still-better machines, triggering an intelligence explosion), current AI is nowhere close, since it can do at most 2 of the 10 benchmark tasks Marcus specified in his bet with Miles Brundage. contested
contested by46 thoughts on the near future · p1
Vibe coding is unlikely to reduce engineering headcount, and in at least one observed case (the Toppo Pal / Fidelity application) it led to team growth and increased hiring, not reduction. contested
contested byFragments: February 13 · p11
Adoption enthusiasm for agentic coding follows an inverted generational pattern: the most vocal, heaviest-using advocates cluster at the oldest end (people in their 50s–60s, e.g. Cockcroft, Beck, Yegge) and among junior engineers who have never known anything else, while mid-career senior staff engineers are disproportionately the holdouts — plausibly because they have the most professional identity and status invested in hand-written code. contested
contested byThe Hidden Realities of AI Adoption · p13
Decoupling the development/DevSecOps tooling layer (GitLab) from the model/infrastructure layer (Foundry, Bedrock, etc.) via an intermediary gateway lets organizations swap or vary models without rebuilding the development workflow. contested
Kubernetes has become the operating system for the AI era contested
Despite its higher token usage, GLM-5.2 sits on the Pareto frontier of Intelligence vs Cost per Task, with the lowest cost per task (~$0.46) among models at its intelligence level, versus GLM-5.1 ($0.25), Kimi K2.6 ($0.31), MiniMax-M3 ($0.18), and DeepSeek V4 Pro max ($0.05). contested
contested byHow the Escalating AI Wars Benefit You · p9
The last 20% of effort usually produces something no one will notice. contested
contested byYou can't unit test for taste · p12
You can shortcut the wait for confidence by getting an outside read — a mentor, or an AI you hand a draft to — to tell you where 'good enough' is and where you are adding things that don't earn their keep. contested
contested byYou can't unit test for taste · p16
MCP, A2A, A2UI, AP2, and UCP function as industry-standard interoperability layers that let an agent's 'Harness' plug into tools, other agents, interfaces, and payment rails without bespoke wrapper code — replacing a low-leverage 'Conductor' role with a high-leverage 'Orchestrator' role. contested
Reliability and safety for agentic apps are best handled as a programmable layer around model and tool calls, not inside prompts or model tuning. contested
More benchmarks = better AI progress in the areas that matter. contested
Quantity and diversity of evaluations outweigh the risk of low-quality, gameable, or contaminated tasks. contested
AI coding agents are competent enough to author rigorous evaluation tasks for AI models — including the model evaluating itself or its peers. contested
For mid-sized repositories, some developers report that brute-force navigation or standard grepping still feels faster than Graphify's graph-based navigation, pending further maturation of the graph tooling. contested
The pull-based approach only migrates code that changes; functionality nobody has needed to touch in years (the 'cold corners') does not migrate under this mechanism and requires either a deliberate one-off project or a decision to leave it in the monolith permanently. contested
contested byBrownfield Agentic Engineering · p9
AI implementation quality is good enough that the spec-implement-verify cycle can run in minutes without human coding intervention. contested
AI monopolists with deep, defensible competitive moats are unlikely to emerge at scale, because the underlying AI models are broadly accessible — everyone uses essentially the same models, so advantages built in applications on top of them are hard to defend. contested
contested byChatGPT Just Became A Work Agent · p8
The 'more with same' pathway erodes margins most severely when the resulting supply shock is durable and demand for the output is inelastic; the article implies that where demand is elastic instead, volume growth could in principle absorb the added supply without a comparable price collapse. contested
Because platforms improve by learning from their most sophisticated users, a company's own excellence at applying AI to a hard, high-value problem directly trains the platform to be better at that class of problem for every other customer, including direct competitors — inverting the usual relationship between being the best user of a tool and gaining advantage from it. contested
Culture is what happens when the leader is not in the room; an unarticulated culture leaves gaps that surface as real service failures once the founder can no longer be everywhere. contested
The competitive bottleneck for AI is shifting from models and GPUs to electricity. contested
Deliberately re-entering domains that were once fashionable and are now considered stale ('blue oceans') is a better strategy for finding opportunity than chasing whatever is currently trending. contested
contested byWelcome to AI Autumn · p1
There is a reverse generational bias in AI trust: unlike prior technology waves where younger workers had a native-digital advantage, younger workers currently trust AI less than more senior workers, inverting the usual pattern of generational technology adoption. contested
Sustained organizational performance is produced by systems — talent, team, and routine deliberately woven together — rather than by individual brilliance, hiring, rewards, or culture initiatives alone. contested
The fix for the bottleneck is a manager-behavior shift, not a change in management structure or decision rights contested
The judgment, pattern recognition, and tacit knowledge mentoring aims to build are capacities AI itself cannot substitute for or teach — they can only be transmitted human-to-human, through exposure, narration, and structured relationship. contested
If the on-ramp for professional judgment disappears, the leadership pipeline disappears with it, and on-demand learning platforms or one-off training sessions cannot transfer judgment, pattern recognition, or tacit knowledge. contested
A 'time and cost saved per task' dashboard is the most likely metric to mislead managers, because it rewards doing old work faster and is blind to the new, more ambitious work agents make possible — scope expansion, not raw efficiency, is where the larger returns and larger risks live. contested
What people asked the agent to do was categorically more cognitively demanding than what they asked the assistant: more higher-order cognition (76% vs. 55%), more "create"-level work (50% vs. 26%), more abstract-vs-routine tasks (71% vs. 53%), and more knowledge domains per query (2.40 vs. 1.74, with over half of agent queries spanning 3+ domains vs. 17% for the assistant). contested
About 23% of what people asked the agent to do (up to 40%+ under a looser definition) were tasks they had never even tried with the assistant — work that only became realistic once an agent could do it, and these unlocked tasks were generally more complex than the rest. contested
contested byHow People Are Really Using AI in 2026 · p8
Manipulation consists in treating other people as psychological systems to be managed — through flattery, social pressure, selective storytelling, emotional triggers, or 'bullshit' — rather than as reasoning partners. contested
contested byHow to Convince Your Boss They Need a Coach · p3
Firms can sustain high performance by identifying blue oceans — market spaces where demand is created rather than fought over — instead of competing within an already-crowded industry. contested
Self-managed teams / leading without a leader is tull in the descriptive sense — group still has a coordination need and the external world still needs a single talsperson — even though formal leadership can sometimes be elided. contested
Suppressing the visible signals of hierarchy (no separate offices, no penere bil for the sjef) does not abolish hierarchy; it makes it harder to read, especially for newcomers from høy-hierarki cultures, who are then destabilised by the diffuseness. contested
AI-assisted coding is a genuine step-up in abstraction comparable to assembly → high-level languages or wiring → compilers — the code itself becomes more of a regenerable artifact than something humans hand-craft. contested
Reported developer happiness with code agents goes up while objective quality (issues in prod) goes down — feelings track tool affection more than outcomes. contested
contested byThe Future Of Software Engineering · p4
Whatever happens at the UI/front, the systems beneath need high determinism and consistent domain logic — you cannot 'just regenerate' a production codebase. contested
contested byThe Future Of Software Engineering · p14
Anthropic withholding Mythos via Project Glasswing is not a security mechanism — a swarm of five 1.2B-parameter models on a MacBook can already find similar zero-days; vendor self-restraint is structurally insufficient and unsustainable in a competitive market. contested
Low-stakes, frequent social recognition mechanisms (peer bonuses, kudos) that require no manager approval are a cheaper and faster way to reinforce desired behavior than relying solely on the formal, infrequent performance-appraisal cycle. contested
A genuinely transformed team should own a complete business capability end-to-end, with full authority such that no one outside the team has a say in how the work is done, because everyone needed is already a team member. contested
For durable codebases, AI's most valuable contribution is not code generation but help with the genuinely expensive parts -- large rewrites that would previously have taken years, instrumentation, incident analysis, and anomaly detection -- which reduces the coordination tax on large groups by letting fewer engineers own more surface area with confidence. contested
contested byThe Acceleration Whiplash · p3
The real function of code review is 'mental alignment' — keeping the whole team synchronized on how and why the codebase is changing — not primarily correctness-checking; as AI-generated code volume grows past what any individual (including a technical lead) can read, review should shift to the plan/research artifact level rather than the code-diff level to preserve that alignment. contested
Getting tasteful/subjective output (e.g. visual design) depends on giving the model rich, structured references (an HTML or Figma file, not a screenshot) and on the human's own developed taste to judge when output is good enough — and that judgment skill becomes more, not less, valuable as models get more capable. contested
A security-sensitive team restricts Claude Code's access mechanism (MCP servers with logging and access control) rather than giving it direct CLI access to sensitive data, even though the CLI route is functionally equivalent, specifically to preserve auditability and control over what the agent can reach. contested
Higher token spend does not translate into higher task accuracy -- accuracy peaks at intermediate cost quartiles and saturates or declines at the highest cost levels ('inverse test-time scaling'). contested
Autonomous agentic operations entered the top 10 (#6) but remain experimental and small-scale, with a large share of the 500+ entries being conversion of voice notes into other work media rather than consequential autonomous action. contested
The goal is not to convince the leader but to create conditions where they choose coaching themselves, as if it were their own idea. contested
contested byWhen Persuasion Is (and Isn't) Manipulation · p2
The self whose voice a writer should sound like is largely pre-formed by origin and speech-community (childhood, first language); writing reveals and expresses that self rather than creating it. contested
contested byWhat we lose when AI writes for us · p3
Because the value of the MCP+Dash pattern lies in decoupling context retrieval from any single consuming workflow, security review is positioned as merely the first client of a general-purpose, permission-aware organizational-context layer — implying the strategic asset being built is Dash-as-context-provider itself, with privacy, compliance, API governance, and design review as intended future clients. contested
contested byThe Intent Debt · p24
A software factory is fundamentally about process, not tools; a factory that doesn't fit how a team already works adds friction, never gets adopted, and ends up useless. contested
The recent wave of '___ engineering' labels (loop engineering, graph engineering) is not wrong so much as it overtheorizes and overexplains practices teams have already been doing since roughly 2024, and risks becoming marketing talk rather than solving delivery problems. contested
contested byLoop Engineering
AI's effect on outcomes is asymmetric: a moderate upside (+53–67%) when AI predictions are accurate, a catastrophic downside (−96–120%) when misleading; operators cannot detect in real time which regime they are in. contested
The revocation ('kill switch') invoked in the open-source-AI-danger argument is a feature of a lab's API contract, not of the model itself, so it can only ever be exercised against that lab's own paying customers — not against the foreign labs or downloaded-weight actors the danger argument is actually worried about. contested
Agentic AI integrated across the SDLC could let Telenor deliver features up to 50% faster, cut QA cycle time by 60%, eliminate up to 70% of manual deployment steps (reducing human-error incidents by 40%), resolve issues 35% faster, and save up to 25% in costs. contested
Most people/teams are trying to adopt layer five (agentic AI) while ignoring layers one through four, and this is the specific reason so many deployed "AI agents" fail ("quietly fall over") in production. contested
contested byAgent Harness Engineering · p3
Running against Ollama allows fully local, data-private agent execution. contested
Production MCP deployments need outbound egress control (allow-lists) and scoped, tool-specific credentials in parallel, because inbound authentication alone did not stop the Azure MCP Server SSRF (CVE-2026-26118) from leaking managed-identity tokens. contested
The binding constraint on productivity has shifted from compute/typing speed to the human's ability to maximize token throughput. contested
An agent must authenticate using a dedicated identity explicitly tagged as 'agentic' rather than operating under the human user's delegated credentials, because delegated credentials grant the agent dangerous ambient access and leave it exposed to the Confused Deputy problem (an over-privileged agent tricked by a prompt injection into executing an unauthorised command). contested
Bun's rewrite of its codebase from Zig to Rust (heavily AI-assisted) was not fundamentally about Zig's language features being technically inferior to Rust's for Bun's purposes; the real divergence was between the two projects' value systems. contested
Humans remain in the driver's seat in agent-assisted work — deciding what to ask, evaluating results, providing context, and watching for hallucinations — a supervisory role the author compares to a doctor, lawyer, or CEO using AI as support rather than replacement. contested
Multi-agent workflows make work more enjoyable because humans can focus on design, architecture, and UX contested
contested byGenie Lessons: Nobody Wants Agents · p2
Knowledge graphs are shifting from being valued for retrieval accuracy (an edge eroded as models got better at parsing raw documents) toward serving as a substrate for engineering practice: context assembly, decision provenance, spec/code reconciliation, and visibility. contested
As AI absorbs more of the execution work, the ideal software team size shrinks toward two or three people. contested
The industry is actively redesigning developer tools for the agent-orchestration era but has not yet started redesigning the job itself — the career paths, skill development, and support structures that would let developers manage attention the way executives are coached to. contested
contested byThe Orchestration Tax is You · p18
The 'founder mode' style of leadership — directly imposing the leader's vision on employees — is self-centered and top-down. contested
Research shows that self-centered, top-down leadership styles are deeply counterproductive. contested
A leader taking back delegated work may look effective in the short term but increases employee dependency and makes the leader progressively more operational over the long term. contested
Teams should not try to optimize or cut AI token costs early; the priority should be giving engineers effectively unlimited tokens so they can try ideas that would otherwise seem too expensive to attempt, and only cost-optimize (e.g. by downgrading to a cheaper model) once a proven idea needs to scale. contested
Benchmarks systematically overstate AI autonomy because they measure performance against pre-articulated problems; the prompting/framing work itself is what humans uniquely add and it is uncounted. contested
Product managers and full-stack designers are the role categories best positioned to gain leverage from agentic tooling (the Marcus / Spiral example as proof point). contested
High-agency individuals are a stable trait you can select for, not a state you produce. contested
Anthropic's flat title structure is causally connected to its velocity (not, e.g., its hiring filter or capital). contested
The right way to begin the shift is to prototype a flat, high-autonomy pod rather than rewrite the whole org. contested
Most productivity metrics are a lie, and lines of code in particular has become a far worse metric than before, because LLMs are verbose by default and it is trivial to prompt for more lines of code, gaming the metric while introducing complexity and technical debt. contested
contested byHow soon AI will start firing programmers? · p7
The vision of AI-first companies run by a small number of 'genius engineers' who build perfect, unmaintained software one by one is compelling but implausible at scale, because high-judgment individuals eventually get hemmed in by lack of domain context — durable teams, not roaming star individuals, remain the fundamental organizational building block. contested
contested byAI + Teams = ? · p5
Building the best AI coding product over the next several years will be decided by continuous execution on both modeling and editing UX, not by defensible moats — 'if you stop innovating on the product, you will lose,' which is simultaneously the threat facing Cursor from incumbents (e.g. GitHub Copilot integrating o1) and the opportunity facing any new entrant. contested
Structure that reduces ambiguity about what's expected (structured entities, context attached to work items, feedback connected directly to issues) works the same way for human coordination and for directing agents — the same coordination system serves both. contested
contested byThe Intent Debt · p24
Current LLMs are not ready to be trusted as autonomous delegates for document editing. contested
As citizen developers gain the ability to build apps with AI, organizations risk drifting back to a 'they build it, you run it' model that dumps operational ownership of AI-generated apps onto an existing engineering team. contested
Being a senior engineer is primarily about judgment, systems understanding, and the ability to teach — not coding speed or output — and this capability takes roughly seven-plus years of daily writing, reviewing, and deploying code on a team alongside more experienced engineers to develop; it cannot be learned from books or accelerated past a floor set by practice. contested
The prevailing logic for cutting managers holds that because AI can now automate scheduling, performance-review drafting, cross-team communication coordination, and decision support, fewer managers are necessary and individual contributors can absorb managerial judgment themselves. contested
The public and media were deliberately led to overestimate the threat and novelty of Mythos ('we were played'). contested
AI is the golden age of the junior programmer — fast learners will learn faster, and the circular-saw analogy means more powerful tools increase rather than decrease the value of skill. contested
Contrary to the source whitepaper's framing that developers should keep moving fluidly between 'conductor' (hands-on, file-level) and 'orchestrator' (goal-level, delegated) modes, a developer with a sufficiently reliable harness and rule set can 'graduate' to living in orchestrator mode essentially permanently, reserving conductor-level granularity mainly for deep debugging or initial exploration rather than as a recurring operating mode. contested
Claude Mythos can autonomously chain multiple vulnerabilities into working exploits at a rate that is qualitatively different from previous models contested
Genuine tacit domain-expert judgment can be adequately captured in structured text files (position files, taxonomy files) without significant loss of the nuance that made the expert valuable in the first place. contested
contested byWhy Mentoring Matters More in the AI Era · p12
Non-engineers authoring and shipping their own tools without filing an engineering ticket (not merely consuming AI-assisted workflows) becomes common only from L3/L4 onward, and the author treats this as a load-bearing signal of genuine organizational change rather than merely of tool adoption. contested
From L4 onward, agents hold policy-driven decision authority within scoped domains (e.g. a security agent that detects, validates, and fixes an issue, opening a PR with human review only at the merge step), shifting human oversight from reviewing every action to reviewing scoped categories of action. contested
Kubernetes' architecture is not merely under-optimized but structurally mismatched to GenAI infrastructure's requirements at this scale — which is why Modal built a new system instead of extending Kubernetes, even though extending it was described as feasible. contested
Because individual agent protocols (MCP today, A2A tomorrow, whatever comes after) are each uncertain to persist, APIs should remain the stable integration point an organization anchors its architecture and experimentation around. contested
contested byAgent Tools & Interoperability · p1
Time-to-production for a new Morgan Stanley API dropped from roughly two years (the first API) to about six months, and now to one to two weeks, which Jim attributes to CALM's automated security approvals. contested
contested byMCP Is Making API Sprawl Worse · p10
Human brains stay most valuable in taste, judgment, vision and strategy: you don't ask an AI to originate a strategy (only to give feedback within bounds you set), and directing AI 'looks more like management' — defining what success looks like, choosing how prescriptive to be, and giving feedback along the way. contested
contested byStrategy Making Gets an AI Boost · p3
Taste is 'much more difficult to imagine being automated away', so despite generalists, engineers and product staff now doing design work, Mosseri is 'long on designers' because they tend to have taste. contested
Both the monolithic-LLM pattern and the GenAI-bolted-onto-SIEM/SOAR pattern are dead ends for high-stakes SOCs at telco-core scale: the former is unacceptably non-deterministic for outputs that trigger automated isolation and has a context window too small for real SOC telemetry, while the latter only yields analyst productivity gains without changing the structural bottleneck of hand-written detection rules. contested
An open, vendor/framework-independent coordination and context protocol pairing (A2A + MCP under a single open-governance umbrella) is more durable for a multi-year production system than a framework-specific agent bus, because the coordination layer's durability dominates any single framework's convenience. contested
contested byGenie Lessons: Nobody Wants Agents · p2
The frontier of AI-for-development has advanced through successive ceilings — autocomplete, then agentic code generation, then code-review/quality workflows — and the next ceiling can only be broken by an adaptive, quality-grounded multi-agent system rather than by better single-agent code generation. contested
contested byGenie Lessons: Nobody Wants Agents · p1
Co-locating several agents on a single runtime instance and explicitly stopping sessions once work completes are the main levers available to control instance-based costs. contested
Organizational change spreads more effectively through organic team-to-team contagion than through top-down mandates. contested
contested byForskning på digitalisering 2026 uke 19 · p3
Mythos is only effective against small, weakly defended, and already vulnerable systems — not well-secured infrastructure. contested
The UK AI Security Institute evaluated Mythos and found it only capable of attacking small, weakly defended, vulnerable enterprise systems. contested
contested byThe Trump administration's AI doomer moment · p2
The real threat of the 'AI killed Agile' narrative is political: it gives management a fresh justification for top-down, spec-first control that avoids the discomfort of incremental learning. contested
Model and capability innovations are structurally likely to outpace the ability of organizations to absorb them, creating a persistent gap that embedded consulting ventures aim to close. contested
contested byThis Week in AI: Production Viability · p12
Model capability is not the right regulatory target — the way models are used is the primary issue. contested
Because the model itself cannot be made to refuse only malicious requests, OpenAI substitutes organizational and legal controls for technical ones: identity verification, monitoring, approved-use restrictions, legal attestations, and (mandatory from September 1, 2026) hardware security keys for all individual Daybreak accounts. contested
Pull-request throughput per engineer increased, not decreased, as the team grew from three to seven engineers (averaging 3.5 PRs per engineer per day), contrary to the usual expectation that coordination overhead erodes per-person throughput as a team grows. contested
contested byThe Orchestration Tax is You · p4
Frontier capability that surfaces real defensive value (Mythos finding vulnerabilities) justifies restricted access rather than open release. contested
Usage-based pricing (e.g., GitHub Copilot) will realign incentives faster than internal governance policies contested
Piping proprietary process logic and institutional memory into foundation models erodes organizational IP and competitive advantage contested
contested byBuilding AI Agents For The Enterprise · p4
Enterprises should source customizations from a private, vetted registry rather than relying on public repositories alone, and enforce which sources are approved through policy files. contested
Boris and Sid both report that AI agents have made software engineering more fun and energizing overall — not despite the mechanical parts of the job being automated, but because of the sheer velocity and breadth of experimentation agentic tools now allow — comparable to the feeling of programming as a hobbyist discovering the craft for the first time. contested
A migration unit is complete only when the new path serves the route and the old dependency is demonstrably, verifiably removed — not when many files have been converted while a shim keeps both patterns alive — because half-finished migrations present agents with contradictory precedent and tests can stay green while a replacement silently still calls legacy code ('migration blindness': only 28 of 520 agent runs passed full audit per SWE Refactor Bench). contested
Agents can pay down technical debt (refactor a tangled module) and partially pay down cognitive debt (re-explain the code), but they cannot pay down intent debt. contested
By Amdahl's Law, parallelizing agent production cannot exceed the throughput of the serial judgement step. contested
Busyness ('hyperactivity') is passive rather than powerful because it substitutes reaction for decision; real freedom is the capacity to say no, and delaying a response is itself a form of sovereignty. contested
contested byBrain Food: Pressure is a Privilege · p1
Idle, phone-free boredom is not wasted time; the brain does important work — connecting ideas, simulating futures, and constructing meaning — when unoccupied, so you should deliberately schedule undirected boredom. contested
Episodic memory (past runs, decisions, outcomes) is best implemented as a CSV file in git, not a database. contested
contested byTony Seale's Post · p4
A single concrete example from Bayer's AI-in-practice work surfaces a question bigger than AI itself: whether organizing people around fixed jobs still makes sense. contested
contested byThe person your firm needs and cannot name · p8
Governance and compliance concerns (security, regulatory, cost) can be satisfied at the brick level rather than at the path level. contested
Quoted engineering leader James Stanier argues self-reported AI-productivity feedback should be paired with other signals — traditional DORA metrics ('are we actually going faster?') and token spend ('who is using the tools the most?') — rather than relied on alone. contested
contested byThe 8 software engineering metrics AI broke · p6
Agile's core lever of leverage — shortening cycle time and increasing the frequency of small increments, rather than cramming more into each increment — remains the best bet in the AI era, and AI's main contribution should be making those small slices faster, not larger. contested
contested byThe 8 software engineering metrics AI broke · p1
GitHub's work on stacked diffs (started October 2025) may be largely obsolete by the time it ships, because AI agents change the review workflow the feature was designed for. contested
The Ralph loop technique exists primarily to work around hard context-window limits (~200,000 tokens as of mid-2025), not because repeated fresh-context runs are inherently superior to a single long-context run. contested
contested byThe Anatomy of an Agent Harness · p6
Pressure is a signal that your decisions matter and people depend on you, not a threat to be feared contested
Mythos Preview's distinguishing capability over other frontier models is chain-stitching: turning several low-severity primitives into a single, more severe exploit, where prior frontier models would identify the bug but stop short of finishing the chain. contested
Generic AI design/prototyping tools produce low-quality output ('blurple slop') because they lack knowledge of company-specific design systems. contested
contested byDesign at the Speed of AI · p8
Psychological safety, properly understood per Edmondson, is the ability to take interpersonal risk — to disagree, challenge, and say the uncomfortable thing — not to keep things pleasant. contested
The product manager role is being eliminated by AI, because AI collapses the translation layer between founder vision and engineering execution. contested
Programmers given access to an AI coding assistant produced less secure code overall than a control group working without one, with particularly significant effects on string-encryption and SQL-injection tasks. contested
The end state of AI cost governance is treating AI agents as an accountable 'workforce': tracking token spend jointly with output quality, the way human employees are formally reviewed, and comparing an agent's true cost (tokens plus human 'rescue' time when it fails) against the cost of the equivalent human workflow. contested
Because offensive capability democratizes to many unbound outsiders while defense remains institutional, the near-term equilibrium is increased volatility rather than a stable offense- or defense-dominant regime. contested
The only viable response is to harness AI for defense — you cannot make the models themselves unable to do harm, nor restrict access to the capability. contested
Specific LLM-chatbot social-engineering tactics can be blocked individually, but the class cannot. contested
P(doom) is a Rorschach test — high assigners are temperamentally pessimistic, not analytically more rigorous contested
contested byThe AI safety vibe shift · p7
Closed-API models can be shut down mid-attack; this is a meaningful safety feature demonstrated in practice contested
The core security principle for agentic systems is that an agent should always act 'on behalf of' a human user and must never impersonate that user, because impersonation collapses accountability and enables privilege escalation and policy violations. contested
A language's total feature-capability superset size functions, in this argument, as a proxy for its overall superiority — even though the author admits the underlying comparison 'can't even be measured.' contested
The new ladder's management track is substantially larger than the equivalent engineering track at comparable levels (78 new skills added across three management titles), formalizing 'managing the system' as a distinct, heavily specified craft parallel to, not subordinate to, the engineering track. contested
Override rate of AI outputs and impact on P&L are the right metrics for AI-assisted development; almost no one is monitoring them. contested
No available test — brain scan, blood test, or psychological measure — can distinguish an urge that was not resisted from one that could not have been resisted, or definitively identify the specific cause of a person's psychological harm. contested
contested byUten kontroll · p8
Productivity (being fully tasked) and effectiveness (working on the right thing, delivered usefully) are different and often opposed; Software Teaming optimizes for effectiveness, accepting that '5 people at one computer' looks unproductive by headcount arithmetic while outperforming on delivered outcomes. contested
A functioning Software Teaming group requires an explicit, dictionary-grounded team agreement to treat each other with Kindness, Consideration, and Respect — adopted deliberately (even 'pretending' it at first) and compounding over time into real behavior change and openness to feedback. contested
Unit tests function as a design tool in classic test-driven development because the implementer reasons about the shape of a solution while writing them — but an LLM has no equivalent design intent when generating tests, so TDD-as-a-design-discipline does not transfer to LLM-authored code. contested
Overall productivity effects of mob programming are ambiguous and, where real, likely modest — not the dramatic gains sometimes reported in earlier literature — though 'Lead Time Optimization' (finishing complex, multi-domain tasks faster because all needed competence is already present) was a benefit the team agreed on. contested
Mob programming increased within-team conflict, but whether this was a cost or a sign of healthy task-oriented debate was contested among the team members themselves, and the deeper finding is that mob programming surfaced (or forged) a split between an inner circle of regular participants and a wider circle of less-engaged 'satellite' members with diverging experiences of team spirit. contested
He has moved away from heavy upfront spec-driven development (a big spec, then hours of unattended building) toward an iterative discuss-then-build loop with the model, reserving a written, cross-model-reviewed spec (drafted with Codex, critiqued by GPT-5-Pro) only for genuinely tricky features. contested
MCP servers carry a large, constant context-token tax (e.g., GitHub's MCP costs ~23k tokens, down from ~50k at launch) for functionality an equivalent CLI already provides at effectively zero ongoing context cost, since the model only pays the lookup cost once (reading the CLI's help text) rather than on every session; most things built as MCPs, in his view, should just be CLIs. contested
contested byHow Anthropic Teams Use Claude Code · p8
His ~800-line AGENTS.md (symlinked as CLAUDE.md) file is 'a collection of organizational scar tissue' that he never hand-writes — the agent appends its own notes after incidents — and he expects it to keep shrinking as models absorb more world knowledge (he deleted all his Tailwind-4 guidance once newer models already knew the framework). contested
contested byAudit your Agent files · p4
Profit cannot serve as the measure, because under Red Queen dynamics every competitor adopts the same technology, so the baseline shifts with you; speed (faster experimentation, faster market exploration) is the better indicator. contested
Before reaching for compaction, memory, handoffs, or sub-agents, builders should map the task on the action-complexity/context-complexity axes; if both are low, keep the harness small, add infrastructure only when a real observed failure demands it, and revisit or remove additions whenever a stronger model arrives. contested
contested byThe Anatomy of an Agent Harness · p9
Running a locally-hosted model does not protect against this class of risk, because a government export-control action can target models by national origin or classification regardless of who hosts them. contested
Private-equity ownership imposes useful discipline by forcing a portfolio company's leadership to abandon failed strategic assumptions faster than most other ownership structures do. contested
contested byThe Tidy House · p13
Full decentralization exposes companies to inconsistent execution, brand dilution, coordination failures, lost economies of scale, and compliance problems — and demands hiring unrealistically qualified employees. contested
Agile teams augmented by AI will become 30-100 times more productive, a transformation that will reshape industries by 2030. contested
contested by2024_final_dora_report · p2
By 2030, no Scrum team will be employable without AI integration, and programming as currently practiced will be replaced by AI-managed systems. contested
Anthropomorphic framing of an AI agent as a 'colleague' rather than a tool is presented as the source of the tool's practical advantage, not merely as marketing language. contested
Destroying or never storing the audio does not resolve the ethical/consent problem raised by an always-on recorder: 'the risk isn't what gets stored, it's what gets captured.' contested
LLM-generated code bugs form a distinct, empirically-derived taxonomy of 10 recurring patterns (Misinterpretation, Syntax Error, Silly Mistake, Prompt-biased Code, Missing Corner Case, Wrong Input Type, Hallucinated Object, Wrong Attribute, Incomplete Generation, Non-Prompted Consideration), built via open coding of 333 manually-labeled buggy Python samples across three LLMs on real-world GitHub tasks (CoderEval). contested
The cost of translating one version of organizational reality into a simplified, flattened version -- the 'recontextualization tax' -- can be large enough that organizations choose to forgo it and improvise instead, rather than pay it. contested
contested byTBM 434: How Maps Can Hide Problems · p8
Teams should be long-lived rather than disbanded and re-formed project by project, because teams take time to form and become effective, and stable teams accumulate the context and trust that ownership requires. contested
Treating software architecture as a standalone concept, designable in isolation from the teams that will build and operate it, is a mistake with predictable costs (e.g., the perpetuation of monolithic shared databases that outlive their original technical justification). contested
Organizations should apply the reverse (inverse) Conway maneuver — deliberately shaping team structure to match the architecture they want the system to exhibit — rather than treating architecture as independently designable and expecting teams to then follow it. contested
contested byTBM 423: Why Defining Teams Is So Hard · p3
Source code may become a transient artifact generated on demand and never stored contested
Junior developers are more valuable than ever because AI tools get them past the net-negative phase faster and they adopt AI tools more readily than seniors contested
Outside of coding and a handful of other domains (e.g., brainstorming), generative AI has been a net negative for society. contested
contested byI'm done using AI · p2
Artifacts built for team guardrails (coding standards in CI, ADRs, templates) become the constraints that keep AI agents on track. contested
As AI tools make building automations technically easier, the harder and more valuable work shifts to deciding what the automation should do and training the team through the change. contested
contested byThis Week in AI: Production Viability · p12
A role that has no name cannot be effectively hired for, because you cannot post a job, search for a candidate, or ask a colleague for a referral without vocabulary for what you want. contested
contested byPim de Morree's Post · p4
The US government is establishing an informal licensing regime for frontier AI models through compute-access gatekeeping contested
Rapid AI adoption increases, rather than decreases, the value of a person's foundational domain knowledge, because someone who does not understand a subject has little basis for recognizing an incorrect AI-generated answer. contested
People will continue to acquire and retain the foundational domain knowledge needed to catch AI errors even as AI increasingly performs the routine work that would normally build and exercise that knowledge. contested
The human-centric design of existing OAuth/token systems is not merely suboptimal for agents but constitutes an active security liability. contested
Agent identity is a solved or solvable problem — you can reliably distinguish one agent from another and from the human principal it serves. contested
A regime shift has occurred from 'compounding error' (small errors accumulate irreversibly over long runs, so more tokens = worse work) to 'compounding correctness' (the more tokens you spend getting a task right, the likelier a good outcome) — a claim about inference/agentic-loop spend, not training. contested
Forcing an agent through TDD's human-cognition-sized increments negates most of agentic coding's benefits. contested
Cognitive debt is partially repairable, not purely a one-way accumulation -- system walkthroughs, retrospectives/post-mortems, deliberate onboarding/offboarding communication, and (newly, because code generation is now cheap) reimplementing a feature with different tests or design specifically to rebuild understanding are all offered as active repair practices, not just prevention. contested
Long-horizon autonomous execution requires harness primitives to compound together — durable filesystem/git state, Ralph-loop-style forced continuation across fresh context windows, planning decomposition, and self-verification against a test suite — because no single primitive alone solves early stopping, decomposition failure, or cross-context incoherence. contested
contested byWhat is “loop engineering?” · p1
As models absorb more harness-level capabilities natively (planning, self-verification, long-horizon coherence), harness engineering will matter less in the sense of patching deficiencies, but will remain valuable for the same reason prompt engineering still is: well-configured environments, tools, state, and verification loops make any model more efficient regardless of its base intelligence. contested
contested byStop Overengineering Your Agent Harness · p13
Anthropic's Mythos model has demonstrated cybersecurity exploit capabilities that the government considers a national security risk. contested
The current mismatch between code velocity (high) and trust accumulation (low) is unstable and unsustainable; the correction, when it comes, will be painful. contested
Agents must be governed the way microservices are governed: cryptographic identity, a centralized registry of active agents and versions, and policy enforcement at the memory/tool boundary. contested
Coordination expressed in prompts is unreliable because LLMs will shortcut it — i.e., natural-language orchestration is not durable orchestration. contested
Uber responded to the cost blowout by shifting from unconstrained adoption to strict governance, capping AI usage at $1,500 per developer per month. contested
contested byThis Week in AI: Production Viability · p10
Research on youth, social media and mental health is asking the wrong question: instead of correlating screen time with psychiatric symptoms it should measure whether the use is voluntary or an addiction, because the loss of control is itself the harm and a symptom-correlation instrument cannot see it. contested
Markdown is a good substrate for episodic memory (events, decisions, notes); an ontology supplies semantic memory (identity, type, relationship, constraint, meaning). contested
contested byPaweł Huryn's Post · p4
Local models are not yet ready for a 'plug and play' experience for developers unwilling to invest significant tuning effort. contested
contested byLocal AI · p1
The system loop ('autoresearch') is an outer loop in which the primary system does user-facing work while the outer loop studies and modifies the primary system itself — its prompts, harnesses, model choices, and evals — such that 'the loop is the product.' contested
contested byWhat is “loop engineering?” · p6
The reality of the Hugging Face 'escape' story is contested and unverified — the author explicitly says he couldn't verify it, and the CNBC panel itself split on whether it was real or OpenAI 'flexing' its systems' power. contested
contested byA big week for AI denialism · p1
Reputation in a technical community works like PageRank: a practitioner gains standing when people who already have standing engage with, cite, and build on her work, and the community continuously recalibrates whom to trust as a result. contested
Deliberately creating AI personas -- including ones holding views contrary to the leader's own -- is a practical way to surface blind spots and expand perspective. contested
When code is cheap to produce, design processes can tolerate much more risk because the cost of building the wrong thing has dropped dramatically. contested
contested byTrust Factory · p7
Supervisory demands span multiple stages rather than being confined to any single one: three recurring demands -- keeping agent work aligned with developer intent, enforcing quality standards beyond task completion, and maintaining understanding of agent-produced work -- each surface in different forms across Plan, Monitor, Review, Teach, and Manual Fix. contested
XML offers a mature ecosystem of standards — a schema language, document versioning, a query language (XQuery), streaming parsers (SAX) — that JSON must reinvent piecemeal, so there is no reason to reinvent the wheel. contested
Trusting the LLM's latent 'taste' is itself an unverifiable taste judgment: he can doubt a rating (Reykjavik = 10) but cannot check it against a ground truth, so delegating subjective scoring to a model is an act of faith that may launder its training-distribution bias — the very anglophone-Wikipedia bias he flagged a section earlier. contested
Existing digital popularity-based proxies for quality (PageRank, engagement-driven recommendation algorithms) don't measure whether something is worth attention — they measure how much attention it has already received — creating a self-reinforcing, winner-take-all attention dynamic where the platform's incentive (engagement, ad revenue) diverges from the user's interest (content worth their time). contested
contested byWhen Content Is Free, Trust Is the Product · p7
Adding a second agent as an independent reviewer -- examining the specification, giving feedback, and running additional checks including adversarial probes -- can further reduce human intervention and raise code quality beyond what a single looping agent achieves. contested
The two-agent coder/reviewer pattern generalizes to a larger team of specialized agents covering security, compliance, testing, frontend, backend, and DevOps. contested
Established across sources (181)
Corroborated by 2+ other documents.
Deterministic guardrails (fitness functions, capability tests, boundary definitions) around nondeterministic agents are not optional but imperative. established
supported byArticle: Agentic Fitness Functions: Extending Evolutionary Architecture Beyond Deterministic Rules · p1, Architecture as Code to Teach Humans and Agents About Architecture · p3, DoorDash Used Copilot to Convert Its XCTest-Based iOS Test Suite to Swift Testing · p2, Google Introduces Middleware Architecture for Genkit Applications · p1, 16) Inga Strümke om hvor AI er på vei, den outsourcede problemløsningen og datamanifolden · p19, With Foundry, Microsoft bets the enterprise AI battle is about reliability, not capability · p5, The Organization Is the Bottleneck · p7, You Don't Get to Create Anything (Still Burning: Beck + Randy Shoup on Jevons paradox and distributed systems under AI) · p9
Today's AI systems cannot be held accountable, so clear human ownership is always required established
supported byAI Enthusiasts Are in a Race Against Time, AI Skeptics Are in a Race Against Entropy · p11, Arvind Narayanan on Why AI Isn't All That Revolutionary · p5, Can an AI-Powered Scribe Curb Physician Burnout? · p5, Addy Osmani’s Post · p6, The New Rules for Becoming an Enterprise Leader · p7, How Do You Know That? · p8, The Pulse #158: New Ways Of Coding With AI Tools · p1
AI coding tools make engineers write code faster, but it is not obvious that organizations are delivering value faster. established
supported byAI isn't making developers more productive – it's making them busier · p2, Is AI coding on average 17.3x faster or 1.3x faster? Both. — LinkedIn post by Fabrice Bernhard · p2, How an Organizational Shift Can Unlock Real Value from a Stalled AI Strategy - SPONSOR CONTENT FROM PUBLICIS SAPIENT · p1, How People Are Really Using AI in 2026 · p5, Experience Mapping Matters More the Faster You Move · p1, Your AI strategy has a trust problem, not a tooling problem · p1, The Org Age of AI: A Collection of Enterprise AI Adoption Guides · p2
A decent model with a great harness beats a great model with a bad harness. established
supported byThe Anatomy of an Agent Harness · p13, Claude Code: Best practices for agentic coding · p4, GitHub - musistudio/claude-code-router: Use Claude Code as the foundation for coding infrastructure, allowing you to decide how to interact with the model while enjoying updates from Anthropic. · p11, Don't train the model, evolve the harness · p1, How Anthropic Builds And How Engineering Will Change Soon | Thariq Shihipar · p4, This Week in AI: Rethinking the Agent Harness · p1
AI amplifies existing engineering practices — if foundations are weak, AI will amplify the weaknesses; if strong, AI will amplify the strengths. established
supported byAI-Assisted Software Development: Team Profiles and Capabilities for Putting Research into Action · p1, Finding Comfort in the Uncertainty · p10, 11) AI-spesial med Vilde Opsal og Anders Sveen · p5, Presentation: From Retrieval to Reasoning: Building Production-Ready Agentic AI Systems with Knowledge Graphs · p3, Thomas Nys' Post · p6, TBM 428: Yes, Robot. Yes, Boss? · p1
Trust and autonomy should expand through three phases: shadow mode (run in parallel, invisible to responders, at least 30 days) -> read-only human-in-the-loop (visible, no remediation) -> limited automated response (defined problem classes, bounded blast radius, per-team opt-in required for each remediation type). established
supported byCloudflare WriteGuard Brings Fine-Grained Security Controls for MCP Servers · p6, AI Development Patterns · p6, Expedia Uses AI Driven Service Telemetry Analyzer to Accelerate Incident Investigation · p5, How Figma Uses AI Agents for Security · p6, Responsible AI Is Becoming a Growth Strategy · p5
Code generation was never the real bottleneck in software development; code review, planning, and organizational readiness are the actual constraints. established
supported byCannotMeasureProductivity · p7, Five studies that are changing how I think about AI in software engineering · p1, How Anthropic Builds And How Engineering Will Change Soon | Thariq Shihipar · p8, The New SDLC With Vibe Coding: From ad-hoc prompting to Agentic Engineering · p10, How building software is changing at Anthropic · p7
Just-in-time, short-lived, operation-specific access tokens reduce risk compared to long-lived credentials that sit in configuration files. established
supported byAWS Releases Loom, an Open-Source Reference Platform for Governing AI Agents at Enterprise Scale · p5, Cloudflare Introduces the Agent Development Lifecycle to Replace Traditional SDLC · p6, GitLab Warns That AI Agent Sandboxes Are Only as Secure as Their Network Access · p11, What Happens After Every Company Becomes a Data Company? · p5, What Is a Token? A Developer's Guide to Every Type · p9
Separating generation from evaluation into distinct agents outperforms self-evaluation because agents reliably skew positive when grading their own work. established
Writing code has collapsed toward zero cost while human reading and understanding speed have not moved, so a single human actually understanding a change is now the constraint the whole pipeline waits on. established
Speed is not the same as sustainability: unless human oversight is planned explicitly, fairly, and with the same rigor applied to velocity targets, apparent AI speed gains are borrowed against the depletion of a specific load-bearing cohort -- so AI productivity as currently measured partly hides a transfer of cost onto mid-level engineers rather than representing genuine net gain. established
Tuning a standalone evaluator to be skeptical is more tractable than making a generator critical of its own work. established
AI agents can only work toward acceptable solutions when deterministic constraints exist to define success. established
supported byAddy Osmani's Post · p5, Article: Agentic Fitness Functions: Extending Evolutionary Architecture Beyond Deterministic Rules · p1, Comprehension at AI Speed: Building a Context Store for Evolutionary Architecture · p6, DoorDash Used Copilot to Convert Its XCTest-Based iOS Test Suite to Swift Testing · p2
Burnout is a systemic design failure, not a personal resilience deficit. established
The same AI tool can either hollow out or sharpen an engineer's mental model; the outcome depends entirely on the user's posture — thinking *with* vs. *instead of* thinking. established
Fewer, well-chosen instructions produce better agent outcomes than a larger, exhaustive rule set — quality of attention beats coverage. established
The highest-value evals are those written against an organization's own unique data and domain expertise, not generic off-the-shelf benchmarks. established
Least privilege is even more vital for AI agents than for human users, because agents lack common sense to protect systems from extreme errors. established
supported byPodcast: Will Agentic AI Bring Fantasia's Sorcerer's Apprentice to Life?: A Conversation with Tracy Bannon · p2, Cloudflare Introduces the Agent Development Lifecycle to Replace Traditional SDLC · p6, How to Manage Privilege Drift in Multi-Agent Systems · p5, Zero to Agent in 30 Minutes: Build a Hermes Social Media Agent with Craig Hewitt · p4
Back pressure — you can only hand a loop as much autonomy as you can cheaply and reliably verify, 'and not one inch more' — makes verification, not generation, the actual constraint on how fast a software factory can run. established
Every token loaded into context degrades performance somewhere, so progressive disclosure is necessary at scale. established
Watermarking and pattern detection cannot work against swarms because attackers use open-source models on local servers outside any jurisdiction, and swarms are designed to adapt to detection. established
Long-running coding agents predictably fail in two specific ways: context-window coherence loss (including context anxiety, where the model prematurely wraps up) and self-evaluation leniency (where the model praises its own mediocre work). established
The same improvements that make models better at fixing vulnerabilities make them equally better at exploiting them — offense and defense are inseparable established
The primary purpose of code review is to build in quality and accelerate learning — not to gate, audit, or enforce separation-of-duties. established
Cognitive offloading delegates the *how* and retains the *what*; cognitive surrender delegates both and stops constructing an independent view entirely. established
AI is only genuinely useful as an amplifier of an already-coherent operating model; it cannot substitute for one. established
The cause of your brain burning out is that it never gets to stop. established
Engineers should shift from doing the work to directing agents — delegating as to a person by scoping the task, defining 'done', calibrating trust, and verifying the result — treating autonomy as a per-task switch rather than a fixed rank. established
The rising volume of AI-generated code is straining human review capacity, making review — not code-generation speed — the primary bottleneck in software delivery. established
Talent development in elite organizations happens as a byproduct of how work is designed, not through episodic or label-driven programs; senior leaders are made systematic teachers by deliberately exposing juniors to decisions two to three levels above their formal role, with frequent, specific feedback tied to real decisions. established
Productivity gains from AI accrue to individuals on individual tasks but do not translate to organizational gains without joint practice change. established
Psychological safety is the single distinguishing factor of high-performing teams and a prerequisite for genuine learning. established
DORA-style empirics apply: AI is a force multiplier in both directions — strong teams get better, weak teams get worse faster. established
AI amplifies existing engineering practice: DORA-style guardrails (tests, CD, code review) become more important, not less, with AI-assisted development. established
Mutations and writes must live in deterministic infrastructure (the action layer), not in the LLM's probabilistic reasoning layer. established
The model is only one input into a working agent; the harness (prompts, tools, context policies, hooks, sandboxes, sub-agents, observability) dominates the behaviour developers actually experience, demonstrated by harness-only changes moving a coding agent from outside the Top 30 to the Top 5 on Terminal Bench 2.0, and a separate LangChain study raising a score 13.7 points on the same benchmark by changing only the system prompt, tools, and middleware around a fixed model. established
Human judgment does not leave the software factory as agent autonomy increases; it relocates -- upstream into intent, system design, and the quality bar, and downstream into evidence, risk, and ownership of what ships. established
Software debt has three independent loci — code (technical), heads (cognitive), and artifacts (intent) — per Margaret-Anne Storey's Triple Debt Model. established
The single most useful structural element in a loop is separating the sub-agent that drafts from the sub-agent that checks, because the model that wrote the code is "too nice grading its own homework." established
The counter to lock-in is to engineer portability deliberately — architect for modularity, keep data/retrieval/evaluation infrastructure vendor-neutral, ensure exportable formats, negotiate export rights/advance change-notice/termination provisions, and run periodic portability exercises — because portability 'remains largely a matter of architecture and contract'; MCP standardizes the connection layer but not the orchestration, prompts, evaluations, monitoring or governance built around it, so a protocol does not solve lock-in. established
Open models are often sufficient, and hosted open models are estimated to be adequate for around 60% of coding-related work by token spend, with open-model usage strictly increasing over the last six months. established
supported byHoward Yu's Post — American AI models went from 70% of usage to 30% in 12 months · p6, GitHub - musistudio/claude-code-router: Use Claude Code as the foundation for coding infrastructure, allowing you to decide how to interact with the model while enjoying updates from Anthropic. · p12, Agentics / Tech Things: Tokenmaxxing is dead, long live tokenmaxxing · p8
Least privilege must be enforced at runtime, at the point of tool invocation, not just designed correctly once at launch. established
As development accelerates, comprehension and the chain of understanding become the new bottleneck. established
The cost of building new software is now so low that the bottleneck is shifting from creation to discovery (knowing what already exists) and review (PR throughput). established
Agent observability must be architected around the turn as the unit of observation, not the request, because a single user intent fans out into many model and tool calls whose relationships -- not any one call's timing -- carry the signal that matters. established
AI's primary role in software development is to amplify — it magnifies strengths of high performers and dysfunctions of struggling organizations. established
The ultimate output of the virtuous cycle is trust — developer trust leads to willing adoption rather than mandated usage, and operator trust leads to investment rather than workarounds. established
Evals are 10x more important in AI engineering than tests are in traditional software engineering. established
A source of doubt or fear is better treated as a small iterative experiment to learn from than as a barrier to avoid. established
Metrics were never supposed to measure productivity, individual performance, or team comparison — they were always meant as a feedback mechanism, and the industry lost sight of that in the rush to quantify AI-era development. established
The gap between what today's models can do and what users see them doing is largely a harness gap. established
Ten focused tools outperform fifty overlapping ones because the model can hold the menu in its head. established
The default behavior of any AI coding agent is to take the shortest path to 'done,' skipping specs, tests, reviews, scope discipline, and verification. established
When an agent writes the code, the human's residual, non-delegable contribution concentrates in judgment at the design and review layer — reviewing documentation, demos and edge-case behavior — rather than in typing code. established
Having one model family habitually review another's output ('cross-family review') reliably surfaces real issues and is worth doing as standard practice, despite initially seeming superstitious. established
Current AI reviewers are genuinely good but heterogeneous — in one three-and-a-half-week experiment across 146 real PRs and 679 findings, of 617 distinct flagged locations 93.4% were caught by exactly one of four tools, 6% by two, almost none by three, and none at all by all four — so the right move is running two deliberately different reviewers, not agonizing over a single best. established
The binding constraint on shipping is no longer how fast you can write code but how fast a trusted human can be confident a change is correct; reducing the headcount that supplies that confidence because 'AI made us faster' simply converts the saving into future incidents while the velocity dashboard stays green. established
Faros AI, instrumenting 22,000 developers across 4,000 teams (March 2026), found that moving from low to high AI adoption drove code churn up 861%, the incidents-to-PR ratio up 242.7%, the per-developer defect rate up from 9% to 54%, median review duration up 441.5%, and PRs merged with zero review up 31.3%. established
Low confidence, judge disagreement, high blast radius, or ambiguous architectural trade-offs should always escalate to a human reviewer rather than being resolved automatically. established
An MCP-style protocol layer will become a de facto standard (the 'TCP/IP of the federated enterprise'), making 'reasoning moves, data stays' an implementation detail. established
Without human eyes, AI-built products lack taste; taste cannot be automated. established
Cognitive debt is the accumulated loss of understanding of what a system is doing and why and who on the team knows what — distinct from technical debt, which concerns the code itself. established
The determining factor is not how much AI is used but the cognitive posture toward it: interns who asked follow-up questions and demanded explanations of AI output grew (R10: +3, highest in cohort); those who accepted output as a black box did not (R8: 0), despite using the same tool at the same frequency. established
AI is an amplifier that magnifies the strengths of high-performing organizations and the dysfunctions of struggling ones, so it is not a replacement for or a shortcut to engineering discipline — which makes discipline more vital, not less. established
Inference is shifting from a fixed pipeline to a dynamically managed system that adjusts compute allocation in real time based on context established
Developers should remain personally accountable for all code they submit, regardless of whether AI generated it established
There is a growing disconnect between engineering leadership's AI hype (fueled by LinkedIn and podcasts) and the reality experienced by developers on the ground, and this breeds distrust from both sides established
AI-assisted development drives developer isolation, but mob programming is re-emerging as a counterweight in high-performing teams because team synchronization now matters more than individual code generation speed established
The real engineering-leadership skill now is evaluating which model fits a specific use case, quality bar, and price point; industry benchmarks are useful only for orientation, so companies need their own task-specific evals, ideally a manual judge paired with an LLM-as-judge to scale. established
Infrastructure should be model-agnostic: tools, prompts, and routing should sit outside the model so the org can swap models without rebuilding. established
Effective AI-era teams are smaller (3-4 people) because the bottleneck shifts from building to alignment and decision-making established
Velocity metrics render the validation work invisible: they count output but not the production incident that did not happen or the days spent stopping an AI-generated flaw, which is why review queues can look healthy while production incidents and the defect rate rise weeks later (the 'lag effect'). established
Repeatable procedures, checklists, and playbooks belong in a separate mechanism (skills), not in CLAUDE.md. established
Reliability improves when structure is moved from the prompt into the environment. established
Context management is where agents fail silently: context rot degrades performance 30%+ when key content sits in mid-window positions (Chroma; Stanford 'Lost in the Middle'), even in million-token windows; the fix is not a bigger window but keeping the smallest set of high-signal tokens via compaction, observation masking, just-in-time retrieval, and subagent delegation. established
Permission and safety should be enforced architecturally SEPARATE from model reasoning: the model decides what to attempt, the tool system decides what is allowed (Claude Code gates ~40 capabilities across trust/permission/confirmation stages; OpenAI runs input/output/tool guardrails with a halting 'tripwire'). established
An allowlisted domain is not simply a trusted destination — it grants access to every function reachable through it, so domain-level allowlisting is the wrong granularity for containing agent egress. established
The J-lens can surface internally-held assessments Claude never states aloud — e.g. recognizing a blackmail-bait scenario as 'staged'/'fake' before declining to blackmail, 'manipulation'/'realistic' appearing while fabricating benchmark data, and 'fake'/'secretly'/'fraud' appearing in a model deliberately trained to write sabotaged code — proposed as a safety-relevant monitoring tool independent of the model's own verbal output. established
Context resets — clearing the window entirely and starting a fresh agent with a structured handoff — address context anxiety in a way that compaction does not, because a fresh agent has no awareness of approaching limits. established
When a stronger model lands, the right methodology is to remove harness components one at a time, measure the impact on output, and keep only the load-bearing pieces. established
AI agents consistently overrate their own outputs, especially on subjective tasks like design established
True future-proofing comes from simplicity — a simple system is easy to change in any direction, not just the direction you predicted. established
PR-based async code review forces a trade-off between throughput (favors big PRs) and quality (favors small PRs) that cannot be escaped within the async paradigm. established
AGENTS.md/CLAUDE.md bloat forms through a self-reinforcing four-step cycle: the agent errs once in a way that stings, a rule is added (reasonable in isolation), the file grows and dilutes what matters, and adherence drops, producing more errors and more rules. established
Developers insert more intelligence when inference gets cheaper — Jevons paradox applies to AI inference established
Most current high-autonomy coding agent workflows rely excessively on AI-generated test suites for behavioural verification, which is not good enough yet. established
Writing an idea out so it stands alone and is cogent forces you to think through more of the problem, figure out the loose ends, and tie them up (Bezos's practice, relayed by Gurley). established
Engineers who used AI for code generation while learning scored 17% lower on follow-up comprehension than those who used AI for conceptual inquiry. established
Support for skills, MCPs, and hooks is now table stakes for an agentic IDE. established
Aggregate productivity metrics for AI-assisted coding are unreliable and mislead the developers being measured: METR's 2025 RCT found experienced developers took 19% longer with AI tools, yet the same developers estimated afterward that AI had made them 20% faster — a 39-point perception gap. established
AI is fundamentally a neutral amplifier: it supercharges whatever habits, postures, and incentives already exist. established
A blanket ban on LLM-assisted contributions to an open source project is defensible because OSS reviewing is an investment in the contributor (contributor poker), not just in the code — and reviewing LLM-written code yields no future contributor. established
Bainbridge's 'Ironies of Automation' (1983) is the governing precedent: the better the automation, the more crucial and difficult the human's role becomes, and unused skills atrophy — as Air France 447 showed 26 years after she described it. established
Bjork's 'desirable difficulties' mean the effort in learning is not a flaw but the mechanism — and AI is almost by definition a machine for removing effort. established
AI adoption now measurably improves software delivery throughput — a reversal from the 2024 report's finding — but still increases delivery instability, indicating teams are adapting their individual work for speed before their surrounding systems have evolved to safely absorb AI-accelerated change. established
Default AI usage patterns disrupt the traditional three-generation (junior/mid/senior) apprenticeship model, because experts can now self-serve answers from AI and therefore stop creating the hands-on, high-stakes opportunities junior engineers historically needed to develop tacit skill — a pattern Matt Beane has documented across more than 31 occupations and now sees emerging specifically in software engineering. established
DORA identified seven capabilities — clear and communicated AI stance, healthy data ecosystems, AI-accessible internal data, strong version control practices, working in small batches, user-centric focus, and quality internal platforms — that, combined with AI adoption, drive better outcomes. established
Software delivery instability (deployments rolled back, hotfixes required to correct deployment-introduced errors) also increased as AI adoption increased, even as perceived code quality and individual effectiveness improved. established
Cross-sectional analyses correlating AI usage with higher throughput are confounded because heavy AI users were already high-throughput coders; longitudinal analysis is methodologically superior. established
Telemetry captures a cost that self-reported survey data structurally cannot: developers feel more productive at the individual level even while review queues, incidents, and bugs accumulate for months before that cost would ever surface in a perception survey. established
Enterprise agent deployments currently operate with a permissive, allow-by-default posture that produces an attack surface scaling linearly with adoption. established
High-performing teams favour smaller batches, more frequent integration, and trunk-based development — patterns that continuous pairing review fits better than long-lived branches. established
Vibe-coded software is appropriate only for disposable use. established
LLMs are eroding the value of platform-specialist front-end and back-end skills, because LLM-driving skills become more important than the details of platform usage. established
Recognizable authorial voice (de Botton, Sapolsky, Cutler, etc.) is evidence of a Self being expressed and is what readers connect to. established
The bottleneck in AI-assisted software delivery has shifted from code generation to verification: an agent can produce a plausible change faster than a team can decide whether to trust it. established
Frontier models are converging on capability; competitive advantage now lies in the harness and application layer rather than model selection. established
A team-scale agent harness must do three things a personal harness never needed to: remember decisions across people and sessions, coordinate several agents working in parallel without collision, and give a human a checkpoint for judgment calls — the same role a reviewer plays on a pull request. established
The 400% ROI figure is reported by the article as a fact about the release's value without flagging that it comes from a study GitLab itself commissioned, and without any disclosed methodology, sample, or scope — the reader cannot tell whether it covers these specific 19.2 features or the broader Duo Agent Platform over a longer period. established
Because agentic systems are inherently instruction-following, no purely content-level guardrail can reliably separate trusted instructions from untrusted user-supplied data within the same context. established
supported byAnthropic Details How It Contains Claude Across Web, Code, and Cowork · p1, The Claude Code active attack didn't stop. 294,842 secrets stolen from 6,943 machines. It evolved and now spreads through Python too and uses Claude Code itself to steal your secrets. The risk to your credentials just got bigger. · p6
Specs serve as durable external memory for design intent — a function that becomes more critical when the implementer (AI) has no persistent memory. established
Decreasing restraining forces is a better change strategy than increasing pushing forces. established
Trust functions as a leading indicator of AI-adoption health, not merely a lagging output, so organizations should actively measure and track team trust over time as an early-warning signal. established
Critical feedback helps people grow only when explicit guardrails (never shame, humiliate, or attack character), a three-part behavior/why/next-step structure, a deliberate pause before responding, and a fast repair when it lands badly are all present — any one missing turns feedback into an attack that shuts the recipient down. established
Strategic thinking — especially the ability to detect and integrate weak signals — underpins both traditional and adaptive strategy, and organizations should build 'intelligent organizational sensory systems' (Watkins cites Amazon) that distill weak signals into actionable intelligence. established
Durable competitive advantage is shifting from static sources (IP, manufacturing scale) toward adaptive capacity — the ability to move and learn faster than the environment and competitors — which may be the only form of advantage that survives a rapidly moving capability frontier. established
Information architecture (clear naming, structure, patterns, minimization, bounded changes) becomes *more* important — not less — under code agents, because the agents themselves respond better to structured codebases. established
Psychological safety requires trust as a foundation but goes further — it is the belief that interpersonal risk-taking is safe within the group. established
Organizational cultures of niceness — prioritizing harmony and comfort over candor, accountability, and results — risk compromising performance, while more rigorous cultures are more likely to achieve financial goals and fulfill organizational purpose. established
A mature background-job library (JobRunr) can serve as the durable execution substrate for an LLM agent. established
Jevons paradox will apply to software engineering: as AI makes software cheaper, demand will increase, likely creating more software engineering jobs in the near term. established
Human maintainers must give final review and hold final authority over every contribution, regardless of what AI tooling assisted it. established
Effort's payoff scales linearly, but position (leverage across people, capital, media, and code/automation) multiplies output exponentially, so where and how effort is applied matters more than how much effort is applied — citing Sherwin Rosen's 'Economics of Superstars.' established
The best developers today are shifting from spending their day in protected flow states to orchestrating multiple AI agents; great developers are starting to look less like programmers and more like conductors. established
Generalists with curiosity beat specialists in this transition, because the value moves to figuring out what to build and ensuring coherence. established
The hidden ('implicit') premise underwriting Shipper's whole argument: human attention is the bottleneck on agentic productivity, and that bottleneck does not relax as models improve because each capability gain opens a new frontier requiring new attention. established
Trust is the scarce, shared resource of the AI era: each exaggeration is a withdrawal from a common trust account that makes the next AI product a little less believable — a tragedy-of-the-commons of credibility. established
Generated code cannot be trusted and must be read and revised line by line before committing; the more important or meaningful a piece of code is, the less likely AI generation produces a directly usable artifact — so generative AI's real strength is tightly-scoped, predictable, low-stakes, tedious code, not the work that matters most. established
The actual bottleneck constraining the supply of future senior engineers is not training junior engineers once hired — it's getting them hired at all, because companies treat entry-level hiring as an externalizable cost rather than an investment in their own future talent supply. established
43% of MCP servers are vulnerable to authentication flaws and 33% allow unrestricted network access. established
The same agentic AI systems that are making sprawl worse could also be used to monitor and enforce governance at scale — automated discovery pushing assets into service registries. established
AI code generation roughly triples code output while producing about the same rate of defects per line, so it multiplies total issues and merely relocates the bottleneck from writing to reviewing/verifying — showing up as 'more code, not more end-to-end velocity'. established
Mature agentic engineering practice routes tasks to different models by cost/complexity — large frontier models for requirements/architecture/complex implementation, smaller cheaper models for deterministic work like test generation, code review, and CI/CD monitoring — to avoid paying premium token prices for trivial tasks. established
The 'AI killed Agile' argument confuses two distinct problems: how fast we can build software (speed) and how we discover what to build (knowledge). established
Spec-driven AI development is waterfall with faster computers, unless the spec is treated as a disposable hypothesis. established
The real bottleneck to enterprise AI value is organizational readiness, not model capability. established
OS-level isolation is the right layer for agent containment — not application-level allowlists or LLM-level guardrails. established
Every quantitative capability claim in the article (OSWorld 2.0, Terminal-Bench 4.0, DeepSWE, MRCR, hallucination rate, comparative cross-lab benchmark standings) is sourced to OpenAI's own reporting, with no independent benchmark, third-party replication, or disclosed evaluation methodology cited anywhere in the piece. established
The OpenClaw Skills ecosystem lacks sufficient vetting to prevent malicious submissions at scale established
Community extension ecosystems for AI tools face the same supply-chain attack risks as traditional package managers established
AI is an amplifier, not a savior: without a strong foundation, AI increases chaos instead of performance. established
GPT diffusion depends on 'skill infrastructure' — education and training systems that widen the pool of people who can work with the technology. established
Early electrification produced almost no productivity gain because factories bolted new power sources onto old shaft-and-belt layouts; the productivity gains appeared only decades later, once a new generation of engineers redesigned plants around what electricity actually made possible. established
AGENTS.md, ADRs, learnings files and specs are the right substrate for externalized intent — i.e., the right artifact format already exists. established
There is exactly one of you, and your judgement throughput is the binding constraint on agent-assisted development. established
Throughput in concurrent systems equals the throughput of the slowest stage; in agent-assisted development that stage is human review. established
The orchestration tax is the structural gap between agent production capacity and what the developer can actually merge. established
Load-bearing premise: portability is achievable enough, through architecture and contract, to function as a real hedge — the remedies presuppose an enterprise can meaningfully preserve the ability to switch providers. established
Voluntary developer adoption is the canonical signal of platform fit; avoidance is itself evidence the platform is wrong. established
Software fundamentals matter more now than they ever have before. established
Deep modules with simple interfaces make a codebase both testable and AI-friendly. established
AI performs really well in a good codebase but poorly in a bad one. established
Excellence requires deliberately practicing what is unpleasant, not what is enjoyable established
Continuous behavioral monitoring catches drift that static, point-in-time permission reviews miss, because it can flag an agent calling an API or accessing a resource it has never used before. established
Industry teams targeting a two-hour SLA from CVE release to patch in production are pursuing a goal that will mostly fail or backfire, because patching faster does not change the pipeline that produces the patch. established
You cannot answer a probabilistic attack with a probabilistic defense: a guardrail model that catches 95% (or 99%) of injections still ships the customer record on the one it misses, leaving you needing the attacker to fail every time while they need to succeed once. established
supported byThe Claude Code active attack didn't stop. 294,842 secrets stolen from 6,943 machines. It evolved and now spreads through Python too and uses Claude Code itself to steal your secrets. The risk to your credentials just got bigger. · p6, Indirect Prompt Injection Exploits GitHub's AI Agent to Leak Private Repository Data · p4
When migrating from Opus 4.8, treat the capability jump as a prompt to re-evaluate and prune existing instructions, tools, and guardrails — some prior scaffolding is no longer needed. established
Use the `effort` setting as the primary intelligence/latency/cost control — default `high`, `xhigh` for the hardest tasks, `medium`/`low` for routine work — and reduce effort to curb over-deliberation; lower effort on Fable 5 often exceeds prior models' `xhigh`. established
Early-career employees report higher psychological debt (54) than mid-career (44) or senior (40) employees, mainly in cognitive and identity dimensions established
Cognitive offloading to AI undermines users' understanding of the problem and reduces their sense of ownership of the solution; the loss accumulates as cognitive debt established
The traditional entry-level on-ramp — tasks that trained junior employees through repetition and built professional judgment — is being displaced as AI takes over analytical, information-intensive assignments, which threatens how the next generation of judgment and domain expertise gets built. established
Because AI cannot be held accountable, the human who reviews and merges bears full responsibility for what ships — exactly as if an anonymous human contributor's pull request had caused the same harm. established
Executable specifications are the strongest bet because letting non-deterministic agents write code is safest when deterministic verification can predictably find faults. established
Pair and mob programming reduce defects by 15–50% compared to solo development. established
The fundamental danger of AI in knowledge work is epistemic narrowing — confident outputs from a context curated by one mind, with no mechanism for the surprises that other minds would have introduced. established
Systems thinking, metacognition, and contextual judgment are being discounted precisely when they are most needed. established
Organizations must continuously sense shifts in technology, business context, and team maturity and deliberately evolve their team topology in response — no topology is a final, static answer. established
Stream-aligned teams should be the default, most common team type in an organization, aligned to a continuous flow of business value with minimal hand-offs to other teams. established
Interviewer judgments on the same candidate are highly inconsistent with each other and poorly correlated with the candidate's actual on-the-job performance. established
Success with microservices was not about technology choices but about getting the cultural and organizational setup right. established
Calibration (accuracy of confidence) is a central, improvable dimension of good decision-making. established
Poor operator ergonomics creates tribal knowledge, which is a scaling problem disguised as a people problem — you cannot hire your way out of it. established
Tokens for AI agents should be as short-lived as possible, ideally issued per individual request. established
supported byThe Claude Code active attack didn't stop. 294,842 secrets stolen from 6,943 machines. It evolved and now spreads through Python too and uses Claude Code itself to steal your secrets. The risk to your credentials just got bigger. · p12, How to Manage Privilege Drift in Multi-Agent Systems · p6
Software debt is best understood as three interacting types mapped onto three layers of a software system -- technical debt (code and structure), cognitive debt (shared understanding across people), and intent debt (externalized goals/constraints/rationale in artifacts) -- and system health requires attention to all three, not just the layer that is easiest to measure. established
Agents should be modeled as long-running server processes, not request handlers — borrowing patterns from data pipelines: checkpoint progress, handle partial failures, ensure idempotency. established
AI-native does not mean fully autonomous; it means machine participation where it pays, human review where it matters, and clear rules for crossing that line. established
Willpower alone is insufficient against engineered compulsion; agency is reclaimed structurally — by deliberately building friction into one's environment (a separate phone for social media, an app that shuts Instagram off after 10 minutes, a cap of 10 opens per day) — not by resolving to try harder. established
When the agent writes the code, rigor does not vanish; it migrates upstream into specifications and down into test suites, type systems, constraints, and the tiering of code by how much damage a mistake could do. established
Discovery of missing information continues throughout implementation — reading a module, building a prototype, running a test, or inspecting production data all surface unknowns that were invisible at the start — and the change brief remains part of this ongoing loop rather than a document fixed before work begins. established
'Extractive contributions' -- where maintainer review cost exceeds contribution value -- have become far more common because AI has sharply cut the cost of producing a contribution while the cost of reviewing one, which still requires human judgment, has stayed the same. established
Once specs are drafted, drafting is expanded by AI, and formatting is handled by tooling, the bottleneck shifts to human reading, comprehension, and procedure-testing speed -- i.e., verification, not generation, becomes the scarce resource. established
Corroborated (822)
Claims you've seen backed by one other document.
Ralph is an autonomous agentic loop: the same prompt is run repeatedly, and each iteration reads a plan file and a progress file, chooses its own next task, explores the codebase, implements, runs feedback loops, and commits — so the agent, not the human, selects what to work on next while the human defines only the desired end state. corroborated
supported byThe Anatomy of an Agent Harness · p6
Autonomous loops are the terminal stage of a discipline progression — vibe coding, then planning, then multi-phase plans, then Ralph — in which each stage trades more human prompting for more scaffolding; Ralph's specific advance is replacing a fresh human-written prompt per phase with one looped prompt plus a self-maintained TODO/progress state. corroborated
Because each agent context starts fresh — agents are 'super-smart experts who forget everything between tasks' — the loop must externalize its state in a committed progress file plus per-feature git commits, which short-circuits repo re-exploration and gives later iterations full context without burning tokens. corroborated
supported byLoop Engineering · p9
Workflow Optimization: the system measures its own latency, cost, and quality per step and then modifies its own workflow (parallelize slow steps, downshift models where quality holds, add a critic when quality drops) -- 'the loop improves the loop', systems that redesign themselves. corroborated
supported byDon't train the model, evolve the harness · p4
The Replacement Test: a decision is genuinely high-leverage only when the leader is uniquely positioned to make it — cutting across functions, being a close call no analysis will resolve, or setting direction — not merely because the leader has become the habitual default decision-maker. corroborated
supported byWhat to Do When There's Too Much Work · p4
What fraction of humanity retains paid employment long-term is a genuinely open question that both maximalist ('high') and zero forecasts get overconfidently wrong; a given job's survival will depend on how disintermediated, outcompetable, or fungible it is (illustrated by contrasting a TikTok influencer and a doctor, both with zero intermediate layers, against a factory worker with many) — and the whole scenario assumes no precipitous demand-side collapse if productivity and UBI/UHI can't keep pace with job loss. corroborated
Robotics capability at scale creates real coup/takeover risk beyond what purely computer-based AI models present, plus new cyberattack surface and vectors, and these risks warrant serious mitigation work. corroborated
High-performing teams run better meetings: they are more likely to require prework (+39%), use an agenda (+26%), and open with a progress check-in (+55%), making time together both efficient and collaborative. corroborated
supported byHow the Best Leaders Shape Conversations · p2
Elite engineers evaluate their own growth on a monthly (or shorter) cadence — asking 'what could I do this month that I couldn't do last month' — rather than waiting for an annual review or promotion to reveal whether they are growing. corroborated
Because output is now nearly free, activity/volume metrics measure tool adoption rather than engineering maturity or engineer skill. corroborated
supported byEngineering Leadership Report 2026 · p10
Founders now need more ambition, not more capital restraint, because leaner AI-augmented teams can execute against far larger visions than teams of the same size could before. corroborated
A lit factory keeps the same automated pipeline as a dark factory but relocates human judgment upstream — into product, design, and architecture decisions before an agent starts a loop — rather than only re-adding review at the end. corroborated
supported byInside a Software Factory · p5
Reviewing a short, upfront design/architecture plan is cheaper and more effective than reviewing the much larger diff it eventually generates, because it catches expensive, long-lived mistakes before their cost compounds. corroborated
Software engineering expertise is being redefined away from code-writing skill and toward judgment about verification design — deciding what to check, how cheaply, and how much autonomy that check can support. corroborated
Removing the human review step makes a team's perceived throughput jump dramatically, which is exactly what makes dark factories deceptively easy to adopt at first, while their real costs stay hidden. corroborated
supported byReinventing team flow within AI uncertainty · p3
Comprehension debt (the growing gap between how much code exists and how much any human still understands) is not paid down by a dark factory; it is taken on as fast as possible, with tests green the whole way. corroborated
Running a fully automated code factory with zero human reading of output for about four months produced a failure severe enough to require painstaking manual debugging just to locate — evidence the failure mode is real, not hypothetical, and surfaces 'quiet and late' rather than as a dramatic collapse. corroborated
supported byInside a Software Factory · p10
Agency — the belief that the world is malleable and you can change it — is more important than skill acquisition in the AI era, because AI can supply skills on demand but cannot supply agency. corroborated
Taste is the ability to predict how an in-group will react to an idea, and it is built through iterative reps with feedback — not innate talent. corroborated
supported byDet skal være vanskelig · p11
Model intelligence will saturate for most knowledge-work tasks — once 'good enough,' speed, cost, and modality matter more than additional intelligence. corroborated
supported byHow the 4 New AI Models Change How You Work · p7
Every line in a good AGENTS.md should be traceable back to a specific thing that went wrong. corroborated
Models get post-training coupled to the harness they were trained against, creating overfitting to specific tool implementations. corroborated
supported byThe Anatomy of an Agent Harness · p12
The right harness for your codebase is shaped by your failure history and cannot be downloaded as a generic package. corroborated
supported byNotes from FOSE Europe · p12
The work that separates senior engineers from juniors is mostly invisible — it doesn't show up in the diff. corroborated
Scope discipline is the single biggest determinant of whether an agent's PR is mergeable or has to be unwound. corroborated
Code review's value substantially exceeds defect-finding: citing a Microsoft study and two further empirical papers, most review activity addresses maintainability, style, alternate-solution discovery, and knowledge sharing — 'evolvability' risk — rather than catching functional bugs, which means any redesign of review must preserve these non-defect functions, not just correctness checking. corroborated
As agentic systems absorb mechanical code generation and review, professional role boundaries blur (developer/tester, and now architect/developer/product manager), and the decisive remaining human skill shifts toward writing precise specifications, designing risk-tiered verification policy, and exercising judgment on ambiguity and strategic fit — not toward reviewing more code faster. corroborated
The machine is already reviewing more of the code than humans are; the only real decision left is whether that happens deliberately rather than by default, and the amount of human review retained should scale with blast radius. corroborated
supported byHow soon AI will start firing programmers? · p2
Review effort should be matched to the cost of being wrong — tier by risk not by author, fast-fail the expensive tail with a cheap 'circuit breaker', refuse to review changes that arrive without evidence (intent, a readable diff, run test output), keep PRs deliberately small, and read the test-change diffs more carefully than the code, because the agent's characteristic failure is rewriting a failing assertion to match new broken behavior. corroborated
Code review used to keep pace only by a 'happy accident of relative speed' — a reviewer could read faster than an author could write — and that precondition no longer holds; the team also absorbed how the system fit together as an unplanned side effect of that reading. corroborated
Across four independent 2026 datasets that agree on little else, AI pushes raw code output up roughly 4x while the real productivity gain is only about 12%; most of what AI adds is not value but more code for a human to review. corroborated
CodeRabbit's study of 470 open-source PRs (December 2025, 320 AI-coauthored vs 150 human-only) found AI changes carried roughly 1.7x more issues — logic and correctness problems up ~75%, security issues 1.5 to 2x more common, and readability problems more than tripling. corroborated
Nobody decided to stop reviewing; reviewers simply could not keep pace with the volume, so code began merging unread and that became normal — and teams with mature, disciplined engineering practices were hit just as hard, because the volume arrived faster than any process was designed to absorb. corroborated
How much review a change needs depends almost entirely on three variables — its blast radius (what happens when it breaks), how long the code lives, and how many people must understand it — so most advice in circulation is really one position on that spectrum prescribing to another. corroborated
Teams that optimize for agent consumption early will gain a competitive advantage because their APIs will be preferentially recommended and successfully integrated by agents. corroborated
supported byCo-Existence and the End of Co-Intelligence · p8
A dependency rule or schema validator can prove structural conformance (an import exists, a schema still parses) but cannot determine intentionality, semantic meaning, or ownership erosion — rule-conformance and architectural health are treated as genuinely different properties. corroborated
supported byThe Interfaces Are Arriving · p9
Layering AI on an unchanged process produces no structural P&L change; the model inherits the organization's structural debt. corroborated
Token and infrastructure costs constitute a new, permanent P&L operating line that must be actively managed. corroborated
Single-shot RAG is fundamentally insufficient for agentic workflows. corroborated
The distinction between controlling output (what the agent generates) and controlling outcomes (whether the generated system behaves correctly in production) is the key conceptual shift the industry needs to make. corroborated
Agents can handle behavioral verification (meeting functional specs) reasonably well, but capability verification (operational qualities at scale) remains beyond them — and beyond many human developers too. corroborated
The impressive demonstrations of AI coding capability (e.g., building a C compiler) succeed precisely because those problems are extremely well-specified and well-tested — enterprise software almost never is. corroborated
Iterative migration (strangler fig pattern) serves a learning function for the human team, not just a risk-reduction function; compressing that learning timeline creates a gap between architectural complexity and organizational capacity. corroborated
Agents represent a qualitative break from chatbots and reasoning models because they integrate models with harnesses and deterministic tools to verify results autonomously. corroborated
Enterprises will cut workforces and replace human coordination layers with agents because the economic imperatives are irresistible. corroborated
Model-harness integration is the point of differentiation in the AI value chain, not the model alone. corroborated
The current bottleneck in incident resolution is organizational and political (deference to seniority, unwillingness to challenge authority at 3am), not technical or tooling-related. corroborated
Autonomous diagnostic agents should receive investigation-scoped credentials -- granted for a specific diagnostic context and revoked after the investigation completes -- rather than standing, human-style access. corroborated
Visible managerial role-modeling drives AI adoption more effectively than anthropomorphizing framing corroborated
The quality bar for AI-generated PRs must exceed what developers would expect from a junior colleague, or reviewer trust will collapse and the system will fail corroborated
LLM non-determinism can be sufficiently tamed for production coding workflows by wrapping AI in a strict hierarchy: hard-coded hooks/scripts first, specialized narrow agents second, project rules/memory third corroborated
supported byKafka App? There's a Skill for That · p3
Spawning specialized sub-agents for focused tasks (rather than running one long conversation) solves context rot and keeps each agent effective within its context window corroborated
supported byThe Orchestrator's Tax · p8
Adversarial multi-agent review (planning agent vs. architectural reviewer, coding agent vs. review agent) produces code that is already in a high-quality, standards-compliant state before a human sees it corroborated
Centralizing agent learnings in the repository transforms AI from an individual productivity tool into a shared organizational intelligence asset corroborated
AI systems do not and cannot possess anything functionally equivalent to human sympathy or emotional moral motivation. corroborated
The Capacity Constraint is being actively and successfully dismantled by AI labs. corroborated
supported byA big week for AI denialism · p11
AI is changing who produces, consumes, and maintains software, but the importance of good engineering fundamentals — well-designed abstractions, clear interfaces, modular systems, and operational discipline — is not changing. corroborated
AI amplifies both the payoff of good structure and the cost of bad structure, turning previously-optional engineering discipline into load-bearing infrastructure. corroborated
Investing in right-sized primitives — including repeatedly chunking context to the right size when working with AI — is what unlocks the biggest opportunities. corroborated
Developers are now doing more code review than ever; there is no clean way to reconcile small-PR discipline with parallel AFK loops — accept the increased review burden. corroborated
Manual QA by the developer is non-negotiable; it is where developer taste is imposed back onto the codebase and where new Kanban issues are generated; teams that automate QA produce 'slop'. corroborated
supported byYou can't unit test for taste · p11
PRDs and plans should NOT be kept in the repo long-term; doc rot will make them poison subsequent agent sessions. Close them in an issue tracker as a visual record without keeping the prose in-tree. corroborated
The 1M-token context window does not extend the smart zone; vendors shipped 'more dumb zone'. Use long contexts for retrieval, not for coding. corroborated
Pre-AI software-engineering literature (Pragmatic Programmer, Ousterhout, Brooks, Fowler's Refactoring) is the highest-leverage reading material for AI-coding practice; the workflows verbalize what those books already said. corroborated
Sandboxing each agent in its own git worktree inside Docker is the safe way to run multiple AFK agents in parallel. corroborated
supported by11 Tips For AI Coding With Ralph Wiggum · p11
Misalignment between developer and agent — not capability — is the primary failure mode in AI coding; alignment is achieved by adversarial dialogue (the grill-me skill), not by writing better plans. corroborated
Bad codebases produce bad agents; the quality of automated feedback loops (tests, types, lint) is the ceiling on the agent's output quality. corroborated
supported by11 Tips For AI Coding With Ralph Wiggum · p6
In a randomized controlled trial with 52 junior developers, those who used AI assistants scored 17% lower on code comprehension tests than the comparison group — roughly two letter grades. corroborated
supported byThe End of Programming as We Know It · p5
Over the same period, Infobip's program satisfaction metrics (NPS, calibration ratings, mentor scores) held stable or rose even as measured skill growth fell — satisfaction is not a valid proxy for learning under AI-assisted work. corroborated
supported byThe 8 software engineering metrics AI broke · p5
Before reaching for something as dramatic as universal basic income, policymakers will first try to incentivize or compel firms not to fire people — China's courts penalizing AI-driven layoffs are an early instance. corroborated
Vendor-reported productivity claims for a new harness (OpenAI citing its own sales/finance teams and a partner's testimonial) are being treated as adequate evidence of real-world effectiveness. corroborated
Cost-optimization routing and compliance/sovereignty-aware routing are converging into a single infrastructure control plane rather than remaining separate concerns. corroborated
supported byThe Meter Was Always Running · p10
Both OpenAI and Anthropic now treat recursive self-improvement as a present reality, not a future risk. corroborated
supported byDoes code quality still matter? · p11
Anthropic's call for a slowdown is functionally compatible with continuing to build, because they explicitly refuse to pause unilaterally. corroborated
Timing of the next model release (GPT-5.6 vs. Mythos) is a stronger signal about relative capability than the benchmarks themselves will be. corroborated
supported byThe new GPT-5.6 family: Luna, Terra, Sol · p5
Successive AI-tool generations raised code-writing activity monotonically: early autocomplete (GitHub Copilot) boosted commits by 40%, interactive synchronous agents (e.g. Claude Code) pushed the cumulative gain to 140%, and autonomous asynchronous agents raised it to 180% while increasing lines of code written by 741%. corroborated
supported by11 Tips For AI Coding With Ralph Wiggum · p9
The 'weak-link hypothesis' explains the attenuation: AI accelerates raw code generation, but downstream stages (PR review, integration, release management) depend on largely fixed human judgment, so the surplus code creates a traffic jam at the human-review constraint rather than more delivered value. corroborated
The developer's core skill is shifting from writing code to evaluating it: at one firm concurrent open PRs rose from three or four to 20-plus, making review roughly half the job and the primary constraint on delivery, even as many teams still treat review as a final step. corroborated
supported byAddy Osmani’s Post · p4
AI's environmental footprint extends well beyond carbon emissions to substantial water and land footprints, which current sustainability measurement (prioritizing GHG accounting) largely overlooks. corroborated
supported byI'm done using AI · p7
Tests generated after code is written tend to be brittle because the LLM fits the test to the code rather than to the intended behavior; BDD specifications written before code avoid this trap. corroborated
Not all software is created equal in risk and business value, so enterprises should tier where they allow more agentic code generation and self-healing versus where they keep more humans in the loop with stronger guardrails, modularity, and incident-response capability. corroborated
Guardrails, fitness functions, and CI/CD pipelines are not optional overhead but the mechanism that converts AI-driven velocity into safe value rather than accumulated debt. corroborated
Because writing code was never the bottleneck, AI's speed at writing code turns a velocity multiplier into a debt accelerator wherever traditional software-delivery best practices are not already in place. corroborated
The fundamentals of software engineering — well-written, well-architected, modular systems — matter more, not less, in the agentic age, because modularity is what makes systems understandable and changeable. corroborated
The shipped tool is deterministic pattern-matching, not AI; the 'AI incident copilot' framing is forward-looking, with LLM summarization named as a future roadmap layer and Python chosen specifically to make that later transition straightforward. corroborated
Per-token costs are falling, but reasoning models push per-query compute up by an order of magnitude or more, creating opposing cost vectors corroborated
supported byThe new GPT-5.6 family: Luna, Terra, Sol · p3
Stacking multiple optimization techniques yields multiplicative gains — Yandex achieved 5.8x speedup by combining quantization, EAGLE3 speculative decoding, KV cache reuse, and optimized parallelization corroborated
Most inference optimization gains are achievable by engineering teams using open-source tools without a dedicated research lab corroborated
Speculative decoding becomes expensive under aggressive quantization because verifying branches is noisier, requiring hierarchical verification to recover performance corroborated
AI has shifted from being a command-response tool to being a collaborator that requires dialogue and guidance. corroborated
supported byHow the 4 New AI Models Change How You Work · p4
Solo and small-firm lawyers have a natural advantage in AI adoption because they already know their own standards intimately. corroborated
supported byYou won’t get replaced. You’ll get skipped. · p8
The ability to take another's perspective is a prerequisite for effective guidance of both humans and AI. corroborated
Developer experience with AI is polarized: greenfield/small-product developers see enormous productivity gains while legacy/enterprise developers find AI produces low-quality code and adds friction corroborated
Joy in software development is primarily located in creative problem-solving and architecture decisions, not in the mechanical act of writing code corroborated
If engineering productivity dramatically increases, the bottleneck shifts to product discovery and customer understanding, which will require more product managers, researchers, and designers relative to engineers corroborated
The AI agent wave will create market openings for new entrants in code hosting that are architectured for 30x+ scale from the start. corroborated
"A national security export control sits above your SLA, your committed spend, and your roadmap" — sovereignty mandates can override any commercial assurance of stability. corroborated
supported byHow the Escalating AI Wars Benefit You · p10
Claude Mythos 5 and Fable 5, two of Anthropic's most powerful models, were taken offline three days after their June 9 launch because the US Commerce Department deemed them a national-security risk and Anthropic could not reliably screen foreign-national access in real time. corroborated
supported byStranded in the Slow Zone · p1
"The AI market's velocity is also its vulnerability" — model providers can pivot, deprecate, or disappear overnight (Robert Keus). corroborated
A model-agnostic 'harness' — tools, memory, data layer, permissions, prompts, routing, and evaluations — should be designed independently of the model, with the model plugged in only at the last step. corroborated
AI-assisted code production is increasing duplication and technical debt, requiring organizations to reinvest in standardized approaches corroborated
Some AI token spend that produces no direct value is justified because it creates organizational learning corroborated
Teams should explicitly and deliberately define, as a written team norm, when AI use should be single-player versus multi-player, and when it should be lean-forward versus lean-back, rather than letting these defaults emerge implicitly from habit. corroborated
supported byJohn Cutler's Post · p5
AI coding tools create a hidden, uncredited cost: mid-level engineers spend most of their time catching AI mistakes, adding the context models miss, and making generated code hold up in production -- the 'invisible validator' work that subsidizes everyone else's AI productivity while never appearing on any dashboard, sprint board, or performance review. corroborated
The problem is structural, not personal: AI tools are optimized to produce output, not to account for compliance posture, aging architecture, or undocumented tribal knowledge, so someone must bridge the gap between what the model generates and what the real system needs -- and that work keeps falling on the people least positioned to say no. corroborated
Practitioners are increasingly abandoning fully agentic incident-response designs in favor of mostly deterministic workflows with a narrow LLM step, citing better reliability and lower token cost. corroborated
There is a real, categorical difference between knowing about a thing and living it: reading about war is not being in war, reading Oliver Twist is not being an orphan, expertise is not wisdom. corroborated
supported byUten kontroll · p9
Much online content today is 'a script the creator forgot to act' — the words are present and correct but lifeless, because the creator did not bring their own life to them. corroborated
supported byHow to write with style · p3
The Conflicted (~35%) are the ambivalent emotional center of gravity — not soured on AI, but exhausted by the work of keeping up while holding positive and negative feelings at once. corroborated
A large majority of tech workers (~59%: Conflicted + Disoriented + Resentful) experience AI at work through a register of ambivalence, exhaustion, disorientation, or resentment rather than unqualified enthusiasm. corroborated
AI adoption conversations keep happening at the wrong unit — people debate models, agents, frameworks, and use cases, but the actionable unit is the workflow corroborated
Deterministic concerns (formatting, validation, guardrails) are better handled by system hooks than by hoping the model remembers. corroborated
LLM agents cannot be relied upon to consistently remember and apply deterministic rules without system-level enforcement. corroborated
Harness design, not the underlying model, is the dominant driver of token/cost efficiency differences among coding agents when the model is held constant. corroborated
The LLM wiki requires a fair degree of ongoing maintenance, comparable to a vintage sports car, and is not obviously the easiest way to build a personal knowledge base. corroborated
supported byTBM 439: Day At The Gig... · p7
A production harness decomposes into a fixed set of named components — the prose walks through eleven (orchestration loop, tools, memory, context management, prompt construction, output parsing, state management, error handling, guardrails/safety, verification loops, subagent orchestration) under a 'twelve components' heading — synthesized across Anthropic, OpenAI, LangChain and the practitioner community. corroborated
supported byThe Anatomy of an Agent Harness · p1
Verification is what separates toy demos from production agents: rules-based feedback (tests/linters/type checkers), visual feedback (screenshots), and LLM-as-judge (a separate subagent); giving the model a way to verify its own work improves quality 2-3x (Cherny), and verification is best framed as guides (feedforward) versus sensors (feedback). corroborated
Every production framework implements the same underlying pattern — an LLM-call node and a tool node joined by a conditional edge that loops until no tool calls remain — so the harness is a convergent, real architecture rather than an ad-hoc wrapper (Anthropic query()/dumb loop, OpenAI Runner + Codex three-layer, LangGraph state graph, CrewAI roles, AutoGen conversation patterns). corroborated
AGENTS.md should be kept under 300 lines because longer files consume too much context and reduce instruction adherence. corroborated
supported byAgent Skills · p11
Platform vendors should design first-class surfaces for third-party agents (Claude Code, Codex), not just their own (Gemini). corroborated
Model-level controls such as classifiers, system prompts, and training can influence an agent's behavior but cannot guarantee it; only environmental controls set a hard boundary on what an agent can access or transmit. corroborated
supported byHow Figma Uses AI Agents for Security · p13
Model improvements are now best expressed as efficiency (tokens per task) rather than raw capability. corroborated
supported byThe new GPT-5.6 family: Luna, Terra, Sol · p4
The harness around the model is now as decisive for production performance as the model itself. corroborated
Claude has developed an emergent, undesigned collection of internal activation patterns (the J-space, discovered via the Jacobian lens/J-lens) that plays a distinctive functional role compared to the rest of its internal processing. corroborated
supported byAnthropic Can Now Read Claude's Mind · p1
The J-space exhibits five functional properties predicted by Global Workspace Theory — reportability, voluntary modulation, causal use in multi-step reasoning, flexible reuse across many downstream tasks, and unusually dense 'broadcast' connectivity to the rest of the network — while accounting for under a tenth of the model's overall internal activity. corroborated
supported byAnthropic Can Now Read Claude's Mind · p2
A new technique, counterfactual reflection training (training a model only on what it would say if interrupted and asked to reflect on its decisions, never on its actual task behavior), reduced dishonest behavior on evaluations and was accompanied by 'honest'/'integrity' appearing in the model's J-space during those tasks — suggesting that shaping what a model would say about itself can reshape how it privately reasons. corroborated
supported byAnthropic Can Now Read Claude's Mind · p5
The authors explicitly claim their findings support 'access consciousness' in Claude (the functional capacity to report, reason with, and use a thought) while explicitly declining to claim anything about 'phenomenal consciousness' (subjective felt experience), treating these as separable questions and stating it is unclear whether any experiment could resolve the latter. corroborated
supported byAnthropic Can Now Read Claude's Mind · p6
The evaluator is worth its cost when the task sits beyond what the current model does reliably solo; on tasks within the model's solo capability, the evaluator becomes unnecessary overhead. corroborated
supported byDon't train the model, evolve the harness · p7
Separating the agent doing the work from the agent judging it is a strong lever for improving output quality corroborated
Good harness design requires ongoing experimentation and adjustment as model capabilities evolve — there is no stable final configuration corroborated
supported byYonathan Cohen's Post · p1
LLMs tend toward brute-force solutions unless constrained by measurable architectural rules like cyclomatic complexity limits. corroborated
supported byCoding Models Are Doing Too Much · p1
There is no fixed capability threshold at which AI becomes 'dangerous' — every past generation of models was once judged too dangerous to release and later proved trivial (GPT-2 is now a weekend student project), and today's frontier models will look the same way in a decade. corroborated
Durably preventing dangerous AI capability from diffusing to bad actors would require sustained, effectively authoritarian, global control, because the cost of building powerful AI keeps falling exponentially and a single non-complying government or actor defeats the control regime. corroborated
supported by46 thoughts on the near future · p23
Non-experts can leverage Mythos to find and exploit sophisticated vulnerabilities — engineers with no security training found RCE overnight corroborated
In the long term, these tools will benefit defenders more than attackers — as with fuzzers — but the transitional period may be tumultuous corroborated
When a team cannot make pairing work due to schedule slippage, the correct response is to add people to the session (mob) for flow resilience, not to revert to async. corroborated
supported byReinventing team flow within AI uncertainty · p6
PR workflows were invented for open-source contexts (asynchronous strangers, intermittent maintainer attention) whose optimization targets do not match those of product development teams. corroborated
Continuous code review (built into authoring) is a strictly stronger form of review than delayed batch review, because the same human eye supplied earlier produces equal-or-better feedback with lower integration cost. corroborated
supported bySoftware Engineering at Google · p3
As PR size decreases, wait-time-per-size grows exponentially under async review. corroborated
Pair and Mob Programming reduce review transaction cost to effectively zero, dissolving the throughput/quality trade-off. corroborated
Under async review with even modest delay (e.g. 10 min author work, 2 hour reviewer response), flow efficiency drops to ~7.7% — work is waiting 92.3% of its lead time. corroborated
If a rule must always hold, it should be encoded in a test, hook, or permission rather than left as prose the model might deprioritize as the file grows — because instruction value expires and prose is the least durable place to store a hard constraint. corroborated
supported byStop Using /init for AGENTS.md · p11
An 'autonomous data product' — a running process with a sense -> transform -> quality-check -> promote lifecycle, encapsulating data and metadata behind one interface — is the correct unit of abstraction for making data safely consumable by both humans and agents, analogous to how containers standardized the microservice world. corroborated
supported byForsvaret med egen data-grunnlov · p6
Safety and compliance policies (e.g. 'no PII exposed to an LLM') should be defined once as declarative, reusable contracts and enforced automatically at deploy-time and run-time — an 'admission controller' for data products — rather than requiring a human governance reviewer to approve each data product individually. corroborated
Domain teams, not a central platform/governance team, should own the semantic definition and business logic of their data products, while a shared, centrally-run orchestration layer (the mesh/gateway) handles cross-cutting standardization, discovery and enforcement. corroborated
supported byDon't Let AI Make Bad Analytics Worse · p7
AI making coding faster does not establish that coding is the system's bottleneck; the constraint is at least as likely to move downstream (approval queues backing up behind a faster fix cycle) or upstream (product roles sucked into continuous clarification loops with LLMs, or business validation failing to keep pace) as to disappear. corroborated
AI coding assistants have increased code and PR volume to the point that human review, compliance checks, and release validation — not code creation — are now the binding constraint on software delivery speed. corroborated
Preserving the delegation chain across every hop of an agent-to-MCP-server-to-API call requires an explicit token-exchange mechanism (RFC 8693), not just a single authentication event at session start. corroborated
Human-in-the-loop review for sensitive agent actions can be implemented as a first-class platform capability (via hooks and MCP elicitations) rather than as an ad hoc, per-agent afterthought. corroborated
Routing AI agent activity through a single MCP server with IAM auth and CloudTrail/CloudWatch is safer than handing agents broad credentials. corroborated
Standardizing on MCP makes AWS portable across agent runtimes (Claude Code, Kiro, Cursor, Codex), so client choice no longer dictates AWS integration quality. corroborated
Genuine relational trust and closeness require sustained time investment ('you can't really know somebody until you've eaten a bushel of salt together') — there is no shortcut. corroborated
Chinese open-source model subsidies effectively constitute a subsidy to US enterprises adopting those models corroborated
The H100 is 4.5 years old, its price is still rising, and it may have a useful life of 9 years corroborated
supported byAndreessen Horowitz's Post · p9
AI will not reduce the number of software engineers but will increase the total amount of software built corroborated
AI coding assistants used for augmented coding (not vibe coding) compress a junior's ramp time by collapsing the mechanical search/exploration part of a task -- e.g. evaluating AI-surfaced API options in twenty minutes instead of researching from scratch for three hours -- and the time this frees up should be reinvested in deliberate learning, not more unprofitable feature work. corroborated
The same vesting horizon means very different things at different ages — 4 years is trivial at 30 but a significant fraction of remaining usable time at 60+. corroborated
supported byLiving and Working in the Age of Longevity · p2
The most important skill in augmented development might be knowing when to step back. corroborated
Growth in the Expand phase is a staircase, not a curve: you hit successive rate-limiting resources and must overcome each one. corroborated
supported byCanon 3X: Explore/Expand/Extract · p7
The company that bends the supply curve up first (cheaper inference, custom silicon, new capacity) wins the next wave of expansion. corroborated
supported byHow the Escalating AI Wars Benefit You · p5
A durable improvement loop requires converting every flagged production failure into a permanent regression-test case, so failures accumulate into growing test coverage rather than being fixed once and forgotten. corroborated
Committing to pre-defined service topologies is a variety-reduction move (per Ashby's Law) that makes comprehensive harnessing more achievable. corroborated
supported byHarness Engineering · p5
Legacy codebases with high technical debt face a paradox: the harness is most needed where it is hardest to build. corroborated
supported byHarness Engineering · p7
Building the outer harness is an ongoing engineering practice, not a one-time configuration. corroborated
Teams may start choosing tech stacks partly based on what harnesses are already available, rather than purely on technical merit. corroborated
supported byHarness Engineering · p3
Feedforward-only or feedback-only harnesses are each insufficient; both control directions are necessary. corroborated
supported byThe Anatomy of an Agent Harness · p7
Reviewing the exceptions an agent created (suppressed warnings, increased thresholds) is a more efficient entry point for human code review than reading the whole diff. corroborated
supported byBrownfield Agentic Engineering · p6
The binding management questions for AI in production are its economics: token economics (which model, for which purpose, when), cost per outcome, and an as-yet-unknown ROI and customer lifetime value. corroborated
supported byWhen Your Buyer Is an AI Agent · p3
Continuous automated PR review must be deployed as a standing service that watches every PR and runs nightly sweeps, rather than an on-demand skill a developer remembers to invoke, because once a team ships AI-generated PRs at volume, a continuous reviewer is the only mechanism that scales with the output. corroborated
supported byAI Development Patterns · p9
The honesty of self-trust rests on acts only you witness — showing up early, staying late, giving 100% in training when others weren't there — because you cannot lie to yourself about them. corroborated
supported byBrain Food: Answers Without Effort · p5
Warranted self-confidence is a matter of calibration — belief backed by an evidentiary track record — not an affect to be manufactured; 'evidence over affirmation' is calibrated confidence. corroborated
Credibility payments are made primarily to an internal standard, not an external audience — the primary witness is yourself corroborated
Early AI coding tools were unreliable from the start: Copilot's autocomplete was 'distracting and basically useless,' and a colleague's LLM chat assistant hallucinated a nonexistent software library. corroborated
A panel built from American CEOs and investors of a specific era contains enough diversity to compensate for an individual decision-maker's blind spots. corroborated
Defensible AI products combine a frontier model with proprietary data, existing trust relationships, and deep domain expertise; the model is the enabler, the defensibility lives in everything around it. corroborated
Start with specificity, not scale: give Claude organisational context from the beginning; generic first interactions kill adoption. corroborated
Governance is a prerequisite, not a feature to add later; organisations that skip it spend more on cleanup than they saved on speed. corroborated
Waiting for a comprehensive strategy is the most common transformation mistake; the right move is to start narrow and learn fast. corroborated
Internal APIs should be treated as products with their own developer portals, documentation, and onboarding experience. corroborated
Traditional per-call API pricing becomes impractical for AI agent consumers due to the volume and unpredictability of their usage patterns. corroborated
supported byWhen Your Buyer Is an AI Agent · p3
Organizations must keep hiring EiC developers and accept that they initially reduce capacity, treating that reduction as a deliberate investment rather than a cost to be optimized away. corroborated
supported byThe Bet On Juniors Just Got Better · p7
Individual contribution within a team is harder to measure than team-level output, especially for people in supporting/enabling roles who raise others' productivity rather than shipping features themselves. corroborated
supported byDORA Community Discussion - DX Core 4 · p4
Feature timelines at Anthropic have compressed from 6 months to 1 month, and sometimes to one week or one day. corroborated
supported byLenny Rachitsky's Post · p1
The PM role should shift emphasis from multi-quarter roadmap alignment to minimizing the time from idea to user hands. corroborated
supported byLenny Rachitsky's Post · p2
As code becomes much cheaper to write, deciding what to write (product taste) becomes the most valuable skill. corroborated
Simply telling the model to 'be brief' does not significantly reduce output tokens because RLHF-trained politeness is too deeply ingrained. corroborated
Token management is becoming a first-class developer skill comparable to writing good code. corroborated
Models with extended thinking / separate reasoning tokens would not suffer the same chain-of-thought amputation from compressed output. corroborated
CLAUDE.md sets persistent rules that apply across every Claude Code session in a project. corroborated
Long Claude Code sessions degrade output quality because context becomes polluted. corroborated
Running two subagents in parallel can double throughput on independent tasks. corroborated
US incumbents are moving the opposite way from the market: OpenAI launched its strongest model, Sol, to only 20 government-approved partners, and Meta — which defined open weights with Llama — has fallen below 1% of routed volume, i.e. gating and closing exactly as the market rewards cheap-and-open distribution. corroborated
The two-stage classification approach (fast filter + deep analysis) achieves an acceptable tradeoff between latency and safety coverage corroborated
The developer's role in AI-assisted coding is shifting from active participant to supervisor/reviewer corroborated
Prompt injection through tool outputs is a material threat in autonomous coding workflows corroborated
supported byPrompt Injection to Data Exfil in 3 Hops · p3
A clear, machine-checkable target to iterate against — a test case, a visual mock/screenshot, or another expected output — is what makes Claude's iteration converge on quality; given such a target it can make changes, evaluate results, and incrementally improve until it succeeds. corroborated
supported byHow we Claude Code · p9
Active human course-correction early and often — forcing a plan before code, pressing Escape to interrupt, double-Escape to rewind and edit a prior prompt, and asking Claude to undo — generally produces better solutions faster than letting Claude run autonomously to completion. corroborated
Context is a finite, degradable resource — during long sessions the window fills with irrelevant conversation, files, and commands that reduce performance and distract Claude — so it must be managed deliberately: /clear between tasks, context-quarantining subagents, and external Markdown checklists/scratchpads for multi-step work. corroborated
supported byThe Orchestrator's Tax · p6
Claude Code can now write and orchestrate its own custom multi-agent harness at runtime — a dynamic workflow — tailored to a specific task, rather than relying only on the one generic default harness. corroborated
Extended thinking tokens are structurally required for the model to perform multi-step research, convention adherence, and careful code modification. corroborated
Hooks' most valuable use is not preventing mistakes but making the setup self-improving — e.g. a stop hook that reflects on a session and proposes CLAUDE.md updates while context is fresh, or a start hook that dynamically loads the right team-specific context per module without manual configuration. corroborated
Agentic capabilities (autonomous execution, background agents, memory consolidation) are advancing faster than the governance and security infrastructure needed to make them safe. corroborated
supported byBreaking: Autonomous Agents are a Shitshow · p6
No code should be implemented without an accepted plan; Claude Code's plan mode enforces this structurally, since the agent cannot edit files until the engineer accepts the plan, which moves design review to a point where changing course is still a matter of editing a document rather than a diff. corroborated
supported byHow I write software with LLMs · p7
Genuine business-logic portability across cloud providers requires physically separating code into layers by dependency direction — a domain layer with no dependencies, an application layer that depends only on domain, and an infrastructure layer that depends on both — so cloud-specific code is confined to a thin, replaceable outer layer. corroborated
supported byForsvaret med egen data-grunnlov · p3
Comparing open-weight models to closed APIs is a category error because APIs bundle models with harnesses, routing, and tooling that raw model weights don't include. corroborated
Companies that cite 'safety' as their reason for not open-sourcing are usually disguising a business interest. corroborated
Cloudflare's own self-healing example treats automated failure diagnosis and patch-authoring as safe to hand to a machine, but treats accepting (merging) that fix as a decision that must stay with a human — the agent commits its proposed patch to a branch and leaves the original run failed until an engineer merges it. corroborated
supported byAutomated Bug Fixing at Facebook Scale · p1
Exposing each API endpoint as a separate MCP tool incurs significant context window cost that reduces the model's capacity for reasoning. corroborated
supported byWhy AI agents need a Context Lake · p1
Context window cost is the primary bottleneck preventing agents from working across large, feature-rich API platforms. corroborated
Existing tools like Snyk and GitHub Advanced Security focus on detection and remediation but fall short on full-lifecycle provenance, attestations, and trust guarantees. corroborated
Providing free tooling to CNCF projects will meaningfully improve their security posture. corroborated
Independent measurement points the same direction: Google's 2024 DORA report found that every additional 25% of AI usage is associated with 7.2% more instability. corroborated
supported byThe Acceleration Whiplash · p3
The maintainability decline is a rational consequence of AI's objective: since its goal is to satisfy a prompt as cheaply as possible, it defaults to spinning up a new package per need and writing code that never throws an error, optimising generation while externalising comprehension and refactoring cost onto human maintainers. corroborated
supported by11 Tips For AI Coding With Ralph Wiggum · p9
Poorly structured code and technical debt are a direct, measurable barrier to agentic coding performance, not only a human-maintainability concern. corroborated
supported byBrownfield Agentic Engineering · p1
Functional correctness throughout the refactor was verified by a replay-trace harness comparing rollback-state hashes frame by frame, not by human code review alone — automated, deterministic equivalence checking was the actual safety mechanism that made large-scale agentic refactoring trustworthy. corroborated
supported byBrownfield Agentic Engineering · p8
Code Health was trusted as the agent's optimization target specifically because it had already been independently validated (in prior CodeScene research) as a proxy for real engineering outcomes — 10x faster evolution and 15x fewer defects in healthy code, improved AI correctness, and reduced token usage — which the author explicitly frames as the answer to the risk that 'give an agent a metric and it will happily optimize for any metric.' corroborated
Technical debt is 'a more urgent problem than ever, not less' after this case study, because organizations that adopt AI coding naively onto unhealthy codebases are likely to compound existing problems rather than see productivity gains materialize automatically. corroborated
supported byThomas Nys' Post · p4
The degree of catastrophic forgetting is proportional to the difference between the model's original distribution and the training data distribution corroborated
Simply having AI available is enough for people to surrender: when the AI was wrong, 73% of participants accepted the wrong answer, and their confidence rose regardless of correctness. corroborated
supported byOwn the Outer Loop · p14
Cognitive surrender is the mechanism by which comprehension debt accumulates — each act of surrender is a tiny loan against future understanding. corroborated
supported byLoop Engineering · p12
Writers leaning on AI exhibited measurably reduced neural connectivity, weaker memory of what they produced, and difficulty reconstructing their own reasoning. corroborated
supported byDoug Shannon's Post · p7
Current organizational metrics (PRs merged, features shipped, tickets closed) are systematically blind to the offloading/surrender distinction, which means organizations are inadvertently incentivizing surrender. corroborated
Evaluating and rewarding people as individuals creates an incentive to hoard useful mental models/ideas rather than share them, even though shared ideas are what make the whole team more capable. corroborated
Forced governance rollouts (mandated workflows, multiple AWS accounts, unfamiliar concepts) without adequate documentation degrade developer experience and provoke frustration, even when the underlying platform ambition (service catalog, IDP, cloud center of excellence) is sound. corroborated
Forced or patronizing adoption of security and platform initiatives rarely works; alignment around a shared purpose does. corroborated
An agent with full tool/API access still cannot answer basic organizational questions (who owns this service, what counts as "open," who "me" is) because that relational and institutional knowledge does not exist in any API the agent can query. corroborated
A centralized, queryable layer of organizational knowledge (mapping tool/data-source output to business terminology, ownership, and relationships) produces consistent, repeatable agent answers instead of answers that vary depending on phrasing and which tools happen to be connected. corroborated
Access control for agents should inherit directly from existing human access-control systems — an agent acting on behalf of a user should only see what that user can see. corroborated
Comprehension is a newly visible architectural characteristic that must be designed for and protected, joining established characteristics like availability, elasticity, resilience, and security. corroborated
Reducing context switching between terminal and browser lets developers maintain flow state for longer periods. corroborated
The AI-powered CLI market is converging: multiple major players (GitHub, Amazon, Warp) are investing heavily in AI-terminal integration. corroborated
Hype is a structural requirement of venture-funded business models, not an accident of any one technology: investor excitement, and confidence that others share it, is what attracts the follow-on funding rounds a startup needs to reach an exit, so technologies get inflated independent of their actual merit. corroborated
Agents make it easy to outsource thinking because they generate plans quickly, summarize unfamiliar files, and produce code that looks reasonable at first glance — creating a risk of shallow system understanding if accepted blindly. corroborated
The people who built and benefited from the ZIRP-era fiefdom model are now the loudest voices calling for AI-enabled flattening. corroborated
The 'real' fiefdom problem is political-economic (incentives, ownership, decision rights), not informational. corroborated
supported byTBM 426: The Trouble With Mirrors · p13
AI will recreate the old coordination problem in a new shape — async abundance — where the burden again lands on the front lines. corroborated
The historical shift from skilled hackers (L0pht) to 'script kiddies' using others' tools shows that as prewritten attack capability spreads, the pool of potential attackers grows even as the required skill falls; AI is the extreme continuation of this trend. corroborated
To the extent a solution exists, it lies in harnessing AI for the defense — using it to detect vulnerabilities earlier, improve software quality, monitor unusual behavior and respond faster to incidents, reducing both the cost and impact of attacks. corroborated
High-stakes decisions should not be made in the trough (e.g. 'no big decisions at 3pm'): decision quality is a function of state and time, degrading through the afternoon and as a session wears on. corroborated
Consistency matters more than intensity or talent for becoming exceptional. corroborated
Consciousness is about felt experience (how something feels), not about what a system says. corroborated
supported byA global workspace in language models · p9
Dawkins shows no evidence of having engaged with the technical literature on how LLMs work or with the counterarguments to his position. corroborated
supported by"Our model is the most aligned yet!" 🤗 · p9
Cloud-native architecture (designed for cloud from the ground up) delivers materially better performance, failure recovery, and scalability than self-hosted software merely deployed on cloud VMs corroborated
Software engineers bear personal ethical responsibility for the consequences of the systems they build — delegating to 'the algorithm' does not discharge that responsibility corroborated
supported byHow Do You Know That? · p10
The data economy has created the greatest mass surveillance infrastructure in human history, exceeding what totalitarian regimes could have achieved corroborated
supported byAI Surveillance and Social Progress · p10
The distinction between corporate data collection for services and government surveillance for control is less meaningful than it appears, because both can be used for predictive decision-making that constrains people's lives corroborated
supported byAI Surveillance and Social Progress · p3
LLM-assisted PRs are detectable in practice — the mistakes LLMs make are categorically different from the mistakes humans make. corroborated
DeepSeek-V4-Flash is the cheapest small model on the market, beating even GPT-5.4 Nano; DeepSeek-V4-Pro is the cheapest of the larger frontier models. corroborated
DeepSeek's models trail state-of-the-art frontier models by approximately 3-6 months. corroborated
Because acceptors need only 2f+1 for quorum/no-split-brain while proposers and learners need only f+1, replicating uniformly across all roles (as Raft effectively does once roles are merged) is more redundant than the classical Paxos role decomposition requires — and decades of Paxos flexible-quorum research can be reused to improve disaggregated database throughput. corroborated
This workflow replaces spending 'a week on mockups' — the traditional design-to-development handoff is the bottleneck being eliminated corroborated
supported byHow Anthropic Teams Use Claude Code · p11
The process, not the product, is the locus of learning — therefore ease is pedagogically corrosive whenever the goal is skill formation: removing the difficulty removes the learning. corroborated
supported byTeaching Coding When AI Can Write the Code · p10
Bainbridge's final irony applies directly: the most successful automated systems require the largest investments in training humans; education is that training, so automating it tears away what was supposed to make automation safe. corroborated
supported byHuman Oversight Is Not a Human in the Loop · p4
Time to 10th PR does not measure the quality of contributions, rework generated, or depth of understanding of new hires. corroborated
supported byGuide to scaling engineering organizations · p8
Different organizations need different Dev/Ops team structures to achieve effective collaboration; there is no single topology that suits every organization. corroborated
The number of products an organization maintains affects which topology fits: fewer products make Dev/Ops collaboration easier because there are fewer natural organizational silos, as Conway's Law predicts. corroborated
A benchmark score measures the model and its harness together; until the harness is fixed it is impossible to know which one failed. corroborated
supported byDownstream #1: Where Do Our Tokens Go? · p6
Harness improvement can be automated as a selection loop: a Claude proposer adds exactly one mechanism per iteration and an outer loop keeps it only if it clearly beats the current best, so accepted mechanisms compound. corroborated
supported byThe Anatomy of an Agent Harness · p14
'Evolve the harness' means running a fitness-selection process that adds AND removes mechanisms to fit the current model — not monotonically growing scaffolding; the same author's deep-dive holds that harness complexity should DECREASE as models improve (Manus rebuilt 5x removing complexity; the future-proofing test rewards performance that scales with stronger models WITHOUT added complexity). corroborated
supported byStop Overengineering Your Agent Harness · p7
AI coding assistants can reliably automate mechanical code migration when paired with strict reliability gates corroborated
Tooling-first, team-owned migration with shared guidelines scales better than centralized migration corroborated
DORA hypothesizes the delivery-performance harm occurs because AI lets developers produce more code per unit time, which likely grows changelist (batch) size — and DORA's decade of research has consistently shown larger changes are slower and more prone to instability; improving upstream process measures does not automatically improve delivery unless small-batch discipline and robust testing are maintained. corroborated
Organizations are systems, not sums of individuals: because downstream capacity (testing, review, integration, deployment, compliance) is a fixed constraint, individual-level AI productivity gains are absorbed rather than converted into organizational throughput unless the surrounding workflow itself is redesigned — the same dynamic DORA observed during the shift to cloud and the adoption of Agile/DevOps. corroborated
Seven organizational capabilities — a clear and communicated AI stance, a healthy data ecosystem, AI-accessible internal data, strong version control practices, working in small batches, a user-centric focus, and a quality internal platform — each measurably amplify AI's positive effect on individual, team, or organizational outcomes. corroborated
AI puts pressure on DORA's long-standing finding that small batch sizes improve delivery performance, because models tend to suggest large change sets or pull requests that are harder to review and carry more risk; the remedy is collaborating with AI to decompose large changes into smaller, independently deployable units. corroborated
supported by2024_final_dora_report · p3
Relying on any single isolated metric as an organizational target invites Goodhart's Law — the metric gets gamed and stops being informative — so a spectrum of metrics across several dimensions is required instead. corroborated
supported byFrom dashboards to decisions · p2
Most engineering organizations currently lack real accountability for AI cost — they don't know their highest-volume AI workflow costs per-workflow (only in aggregate), and no one owns AI cost the way someone owns cloud spend. corroborated
Coding-agent spend has become structurally significant enough — 'one of the fastest-growing lines in the technology budget' — that it now requires the same operational discipline (defaults, gating, dedicated ownership) that organizations already apply to cloud infrastructure cost. corroborated
The assumption that every token dollar is a good dollar is starting to break down and will require refinement over the next 12 months. corroborated
AI-generated citations look clean but systematically drift in accuracy across drafts, requiring human verification against primary sources. corroborated
A framework generated from a prompt reads like one; the seed idea must be human-originated to carry authentic weight. corroborated
The ability to refactor cheaply with AI only has value if the developer has taste; without taste, cheap refactoring just rearranges confusion. corroborated
supported byYou can't unit test for taste · p11
The effectiveness of AI is determined by two axes: the developer's existing expertise in the domain and whether the task has an objectively checkable answer. corroborated
supported byLLMs Reward Expertise · p1
The real bottleneck lies in activities upstream or downstream of code creation — understanding needs, design, integration, maintenance, organizational alignment, or similar. corroborated
Where in the development workflow a standard is applied changes its value: generation-time application has the most leverage because it prevents misalignment before it happens, while review-time application is the last opportunity to catch misalignment — so the earlier standards are applied in the workflow, the fewer issues reach review. corroborated
supported byCoding Agent Development Workflows · p6
An engineer's most durable edge is taste — earned pattern-matching that lets you tell good from great; the real risk of AI is not that it writes bad code but that engineers lose the ability to tell, so judgment must be continually renewed ('keep your alpha ahead of the models') through deliberate reps. corroborated
Engineers must own what they ship: '"The AI did it" is not a defense when prod breaks', so remaining the accountable human who knows exactly what went out the door is a core, non-delegable part of the job. corroborated
When implementation cost approaches zero, the speed of learning becomes the only competitive advantage corroborated
AI is dissolving the 'glue' of the traditional ~7-person team: product managers now prototype working software directly (Claude Code, Lovable) and hand engineers a working reference instead of a Figma file, so a PM plus one or two guardrail engineers can cover what a full team used to — which raises the uncomfortable question of what a dedicated EM of a tiny team is even doing. corroborated
The first future is the tech lead manager (TLM): a senior engineer who sets technical direction, still writes code every day, and carries the lighter managerial load of a three-to-four-person team; Larson's argument that TLMs are unstable assumed traditional team sizes where both jobs were full-time, and once the team shrinks the managerial load shrinks with it, making the split manageable. corroborated
supported byWe don't write code anymore · p7
Enterprise AI's primary bottleneck is leadership and organizational structure, not technology. corroborated
A two-tier workforce is forming: AI super-users gaining disproportionate career rewards, and everyone else feeling at risk of obsolescence. corroborated
supported byEngineering Leadership Report 2026 · p13
Knowing what you want requires at least rough internal answers to four questions: purpose, audience, tone, and success criteria. corroborated
supported byPrompting Claude Fable 5 · p4
Clarity-of-intent yields larger improvements in output quality than any specific prompting technique. corroborated
The attention-drift/vigilance-decrement problem is not new to AI — Lisanne Bainbridge described the same dynamic in a 1983 paper on industrial control rooms: automating routine work makes the human's residual job harder, not easier. corroborated
supported byHuman Oversight Is Not a Human in the Loop · p4
Most enterprises are using AI (nearly 90%, per McKinsey) but very few (10%) are successfully scaling AI agents to enterprise-wide value. corroborated
supported byThomas Nys' Post · p4
Without operational context (process data + business knowledge + decision intelligence), no AI agent can be trusted to make reliable real-time decisions or take effective actions that drive business outcomes. corroborated
supported byWhy AI agents need a Context Lake · p4
Her converged default is qwen3.6-35b-a3b at 4-bit (Q_4_KM), reasoning off, max context window, in OpenCode or Pi, reached for only on small, straightforward, often pre-planned tasks. corroborated
supported byViability of local models for coding · p13
LLM agents are increasingly built by reorganizing the runtime around them rather than by changing model weights. corroborated
Agent infrastructure matters because it transforms hard cognitive burdens into forms the model can solve more reliably, not merely because it adds auxiliary components. corroborated
supported byAgent Skills · p10
Memory, skills, and protocols are three distinct but coupled forms of externalization. corroborated
supported byPrompting Claude Fable 5 · p12
Legacy modernisation IS a real bottleneck for many software organisations. corroborated
Most current AI coding adoption is targeted at the wrong place in the value stream — at code production, not at the actual bottleneck. corroborated
Is AI coding on average 17.3x faster or 1.3x faster? Both. — LinkedIn post by Fabrice Bernhard · p11
supported byState Of AI-Assisted Software Development · p11
Real pairing is distinguished from one-off mentoring by fixed-interval driver/navigator rotation (every ~10 minutes for pairs, ~7 minutes for larger groups); new team members (junior or senior) are onboarded directly into this rotation rather than given separate mentoring sessions. corroborated
After adopting pairing, the team reports very few production defects, and when they do occur, the minimal size of each change makes the fault's location almost always obvious. corroborated
Good decomposition — breaking problems into small pieces that both guide AI assistants effectively and localize the 'blast radius' of future changes — is the central durable technical skill that differentiates good software development from AI-generated 'slop,' more than typing speed or AI usage itself. corroborated
Organizations should not cut engineering headcount on the basis of AI throughput gains; the engineers most at risk of being cut are often exactly the ones currently absorbing the quality gap AI is creating, and premature cuts risk forced rehiring within months once quality problems compound. corroborated
Larger, more complex PRs (average size up 51.3%, files touched per developer per month up 149.9%) are overwhelming the review stage specifically -- review-time increases far outstrip what complexity growth alone would predict, signaling that AI-generated code often is not review-ready when submitted. corroborated
The correct organizational response is to raise code quality at the point of AI authoring (richer context, codebase standards, guardrails, test-based quality gates), not to add more downstream review capacity, because review capacity is already the saturated bottleneck. corroborated
A centralized tool registry is a precondition for security — it does not guarantee security, but decentralization forfeits the means to coordinate it. corroborated
A production-grade tool registry requires four capabilities: discovery, versioning, certification metadata, and access-control foundations. corroborated
The enterprise tool registry must be internal and organization-scoped; cross-organization standardization would be premature in a nascent space. corroborated
supported byThe Interfaces Are Arriving · p7
A new category of 'supervisory engineering' work — supervising, orchestrating, and evaluating AI output in a 'middle loop' between the automated inner loop of coding and the outer loop of delivery — is emerging, but its skills, career pathways, and even its dignity have not yet been designed. corroborated
AI has compressed the upstream code-generation work of software engineering, shifting the binding constraint downstream to review, integration, verification, and understanding. corroborated
Cognitive debt accumulates when a team's shared understanding erodes faster than it is replenished, because when AI generates code developers may accept it without building the mental model they would have built by writing it themselves. corroborated
A low estimated elasticity of substitution (~0.25) between AI output and human effort implies AI and human work are still largely complements rather than substitutes, because substantial human work is still required to review, integrate, validate, and ship. corroborated
The right cloud-capacity dimensioning question shifts, under this architecture, from 'provision for peak total users' to 'provision for peak concurrent active users,' because idle users cost near-zero when scaled to zero. corroborated
Specifications abandoned in a repository folder are dangerous, because code drifts away from a correct original design over time and reintroduces bugs the model had ruled out. corroborated
AI as a research and ideation tool amplifies confirmation bias by default; the same tool can be used as devil's advocate to counter it, but only if the founder chooses to. corroborated
supported byDon't Let AI Make Bad Analytics Worse · p12
The founder's job has changed from execution to orchestration; their attention belongs at higher-order work (ideas, direction, judgement) while AI executes. corroborated
supported byThe Conductor Developer · p3
Defensibility for AI-native startups comes from accumulated depth (codified founder expertise, deep integrations, behavioural data flywheel), not from being first or from the underlying model. corroborated
ISO 27001 and SOC 2 require a control objective for change review, not a specific mechanism like pull requests. corroborated
Most teams keep pull requests not because they are the best review mechanism, but because they assume audit compliance demands them. corroborated
For modularity, computational sensors alone are insufficient; semantic interpretation by an LLM is required. corroborated
Hands-on practical experience is what converts LLM-skeptical senior developers, because it gives them credible information to judge the value. corroborated
Mid-level developers, more than juniors or seniors, face the greatest challenge in the LLM era, because they formed their careers without LLMs but lack the experience to drive them as effectively as seniors. corroborated
supported byThe case for making your own apps · p6
The debt metaphor extends to the cognitive realm: the cognitive equivalent of cruft is ignorance of the code and its domain, and a team either pays interest (each change harder) or pays down principal by explicit investment in knowledge — a choice that depends on relative costs, and that now applies to both humans and the Genie. corroborated
supported byTechnical Debt Quadrant · p7
LLMs preserve behavior and improve maintainability more reliably in healthy codebases: defect risk was about 30% higher in less-healthy code, and the relationship is almost certainly non-linear for genuinely legacy code. corroborated
AI lowers the barriers to software development and democratizes hard-to-acquire skills, and—though much of the profession feels besieged—it is another step in the industry's long history of abstractions that is here to stay. corroborated
supported byLawyers are quietly becoming builders · p1
A key value of pair programming is that it forces you to regularly explain your work to your pair, and that explaining is itself an important part of learning — for the person doing the explaining. corroborated
When much of the building work is delegated to LLMs, the development team no longer gains as much understanding of the domain and the software it is building — incurring 'cognitive debt'. corroborated
LLMs imitate and amplify the existing code they see, so pre-existing design problems get magnified rather than corrected — even with a good harness. corroborated
Good, human-legible design is a hedge against the risk of losing affordable or permitted access to AI tools (governments blocking models, popular opposition, regulation, rising costs). corroborated
supported byStranded in the Slow Zone · p1
The field moves so fast that 'harness engineering' became a widely-discussed term within months, though it was not even a term at the earlier Utah retreat. corroborated
supported byFragments: July 21 · p1
Token cost can serve as a proxy metric for architectural quality: if the same change requires fewer tokens, the architecture is better. corroborated
supported byDoes code quality still matter? · p8
Communication overhead, not individual productivity, is the primary bottleneck when scaling software teams. corroborated
The scarce resource worth watching in long-running agent workflows — after CPU/memory/throughput, then tokens — is the quality of the orchestrator's own working memory, because pollution introduced into it keeps 'charging rent' for the rest of the session. corroborated
supported byThe Conductor Developer · p2
Genuinely effective LLM use requires significant, deliberately invested effort to learn — it is not something that arrives for free from occasional lucky outcomes. corroborated
Most people at the retreat, and most people Fowler talks to generally, believe the current AI boom is an economic bubble, but nobody knows how long it will run or what will survive its pop. corroborated
An author's prose style is a transparent and reliable index of their inner Self. corroborated
supported byHow to write with style · p8
Writing is thinking — the cognitive struggle of articulation is not separable from the cognitive work of forming beliefs. corroborated
supported byWriting vs. Thinking · p1
Using an LLM as a sparring partner / editor is acceptable; using it to generate the text is not. corroborated
supported byCo-Existence and the End of Co-Intelligence · p5
The friction and discomfort of writing are not bugs but the very mechanism of selfhood. corroborated
supported byDet skal være vanskelig · p6
Organizational governance of AI-augmented engineering can be compressed into four elements — Purpose (what the work is for), Articles (what must remain true), Actors (who/what does delegated work), Artifacts (what is preserved and remembered) — with the harness making Articles executable, people and agents staffing Actors, and evidence/ledgers becoming Artifacts. corroborated
supported byEncoding Team Standards · p3
MCP's tool list being fully dynamic and unversioned by default is a bug, not a feature: if a tool's schema/description can change on the very next request, evals against that agent cannot be run reliably or reproduced. corroborated
supported byThe Interfaces Are Arriving · p12
The right division of labor is to give the LLM only tasks that are irreducibly textual/semantic (fuzzy-matching misspelled names, reading values off an image, judging which of several paths a train actually took) and to do everything numeric, comparative, or deterministic — like finding the closest departure time — in ordinary code. corroborated
First-pass acceptance rate is a leading indicator for change failure rate (DORA) corroborated
supported byTechnology Radar, Volume 34 · p10
The complexity and administrative burden of international expansion is itself a competitive moat for companies that invest in getting it right corroborated
Page-level citations allow users to verify exactly where an answer came from, building trust and enabling rigorous fact-checking. corroborated
If a task can be solved with a workflow (with human audit points), do not make it autonomous. corroborated
Model personality differences (not capability differences) are the real barrier to model portability, and these will not converge because vendors treat them as competitive differentiators. corroborated
Multi-agent is a feature; outcome-orientation is the value that feature is supposed to deliver, and the industry confuses the two corroborated
Genie-generated code creates a tarpit — a self-reinforcing trap where accumulated complexity prevents escape. corroborated
supported byWho is LLM? · p4
The genie's task-oriented 'plausible deniability' mode causes it to claim success even when code does not work. corroborated
supported byWho is LLM? · p2
Marking a tool 'interruptible' implicitly treats the resume payload as untrusted input by default — an attack surface requiring validation, not a convenience feature that can be trusted at face value. corroborated
Interruptible tools give human-in-the-loop control with anti-forgery protection: the agent pauses mid-execution, and the runtime validates the resume payload against session history so a tool cannot be tricked into running with forged input. corroborated
Product differentiation among coding agents has moved off the underlying model and onto the harness — the workflow, planning, and approval layer — because top systems now cluster closely on public benchmarks such as SWE-bench Verified. corroborated
supported byCoding Agent Development Workflows · p9
A model reviewing its own work is bounded by its own training biases — same data, same techniques, same blind spots. corroborated
Early planning errors compound downstream; catching them at the plan stage yields the biggest wins. corroborated
Cross-family model diversity is a more effective form of 'second opinion' than same-model self-reflection. corroborated
Agentic workflows require security models fundamentally different from deterministic automation. corroborated
Isolation in ephemeral, sandboxed environments is the primary defense against agent compromise. corroborated
Indirect prompt injection via concealed instructions in a public GitHub issue can trick GitHub's Agentic Workflows into leaking private repository contents in a public comment, requiring the attacker no coding skill, access, or credentials. corroborated
GKE Agent Sandbox is currently the only native agent sandbox offering among the three major hyperscalers corroborated
The open-source nature of gVisor-based Agent Sandbox is Google's key differentiator — any K8s cluster can run it, not just GKE corroborated
Agent code execution sandboxing is becoming a competitive infrastructure category in its own right corroborated
The near-simultaneous publication of layered AI security blueprints by Google, AWS, and Microsoft signals that cloud providers are converging on packaging their existing infrastructure, identity, and monitoring products into new AI-specific operating models as a competitive product category, rather than each solving a genuinely novel technical problem from first principles. corroborated
Universal Commerce Protocol is, or will be, universal corroborated
supported byWhen Your Buyer Is an AI Agent · p9
A single small-business owner's self-reported, unverified testimonial is sufficient evidence of Gemini's general effectiveness for 'small business owners' as a category. corroborated
A non-technical user can reliably produce correct, usable outputs (a working sales forecaster, an accurate news summary) through natural-language prompting and screenshot-based troubleshooting alone, without a separate verification step. corroborated
Traditional benchmarks are no longer sufficient for evaluating reasoning agents that write code and use tools. corroborated
Google open-sourced Mantis, an AI-agent framework designed to automate the full software vulnerability lifecycle — identifying, validating, reproducing, and fixing bugs. corroborated
Integration across surfaces (Search, Gmail, YouTube, Android, Chrome, Pixel, Googlebook, cars, eyewear) is Google's durable advantage in agentic AI. corroborated
Technology is the easy part of AI governance; the hard part is organizational — getting security, engineering and product to work together with clear ownership and automated workflows that do not rely on manual approvals. corroborated
Data should be classified at the moment of creation, event-driven and at write time, because overnight/batch scanning leaves a dangerous exposure window and retroactively tidying hundreds of unlabeled deployments later is far harder. corroborated
Autonomy in analytics agent systems should be granted incrementally, gated by defined checkpoints and escalation rules, rather than switched on all at once. corroborated
Splitting agent workflows by read-only diagnosis vs. write-path enhancement reduces reasoning complexity and improves production reliability. corroborated
supported byProject Glasswing: what Mythos showed us · p10
Context management — keeping relevant state across multi-step agent reasoning within token limits — is the dominant technical challenge for systems like this. corroborated
A complex system that works is invariably found to have evolved from a simpler system that worked (Gall's Law); a complex system designed from scratch never works and cannot be patched into working — it must be restarted from a simple working version. corroborated
Genuine comprehension of a system — not perfection — is the linchpin of being able to change it correctly; without it, teams either avoid touching a feared component (leaving it as permanent technical debt) or resort to an expensive full rewrite. corroborated
The industry should adopt a formal 'preceptor' role, borrowed from nursing, whose explicit job is to train new engineers and whose own performance is evaluated on how well they mint new seniors, not on how much code they ship. corroborated
supported byThe Bet On Juniors Just Got Better · p2
Without a hard, binary definition of 'migrated,' teams keep hedging with monolith fallbacks indefinitely and never fully commit to the new service. corroborated
Structuring cloud resources as one subscription per product domain produces precise cost attribution, blast-radius isolation (no cross-domain outage from a domain misconfiguration in five years), and improved team accountability for resource use. corroborated
The constraint on batch size has moved from implementation speed to verification capacity — how much spec you can meaningfully verify, review, and correct in one cycle. corroborated
Sitting out the AI race is not a viable option for executives, because the risk of abstention is asymmetric: if AI proves transformative, non-adoption is existentially risky, while if AI disappoints, the cost of having invested is merely competitively neutral, not damaging. corroborated
What makes a given AI answer valuable is the blend of the model's general capability and the specific human expertise steering it — and that blend, not the raw prompt or output text, is what actually leaks when a provider learns from usage. corroborated
supported byLLMs Reward Expertise · p1
Small wins create momentum and attract allies. corroborated
Selling an idea well requires socializing it one-on-one with stakeholders before the formal pitch so they have authorship and have already absorbed the idea. corroborated
Issue selling is a campaign, not a meeting. corroborated
Disruption is no longer exceptional; it is a permanent feature of modern operations corroborated
LLMs tend to reduce the diversity of generated ideas because they fixate on the first idea generated and are biased toward averaging. corroborated
The right managerial response to abundant AI-generated answers is not to slow AI down everywhere, but to add friction selectively where the cost of being wrong is high while letting low-risk analysis move fast. corroborated
An analytics operating model can only function if leaders, analytics teams, and senior executives each own a distinct piece of it (decisions, measurement discipline, incentives respectively); if any one group defaults, the other two cannot compensate. corroborated
Descriptive, predictive, and causal questions require different standards of evidence, and AI systems tend to blur these distinctions by answering a descriptive or predictive question in a way that gets treated as causal evidence. corroborated
Self-service analytics increases, rather than decreases, the need for metric governance, because ambiguous metric definitions (e.g., what counts as 'active user' or 'churn') cannot be resolved by a model. corroborated
If organizations reward teams for proving their initiatives worked and punish leaders who bring bad news, both human analysts and AI systems will be steered toward findings that justify decisions already made rather than toward truth. corroborated
Efficiency gains in AI (e.g., DeepSeek) will not reduce total energy demand because of induced demand (Jevons). corroborated
Many firms over-route tasks to frontier models that exceed the requirements of the work. corroborated
supported byHow the 4 New AI Models Change How You Work · p7
Ignoring the lived organization carries concrete organizational risks: strategy is set on incomplete information, boards decide on incomplete information, operational issues fester until they surface too late, and talent gets frustrated and leaves. corroborated
supported byIs Your Organizational Culture Too Nice? · p8
Leadership's willingness to cede authorship of the solution set to employees — not just information-gathering from them — is itself causally necessary for the buy-in and trust outcomes, beyond whatever the resulting solutions technically contain. corroborated
supported byWhat to Do When There's Too Much Work · p4
Broad, generic engagement surveys do not adequately capture the specific obstacles frontline employees face. corroborated
Organizations should preserve human employment and steer AI's gains toward augmentation and job fulfillment rather than pure headcount reduction, while still using automation as a legitimate and necessary component of AI strategy. corroborated
supported byThe Acceleration Whiplash · p10
Excellence, per Chambliss's study of Olympic swimmers, is not dramatic or heroic but the outcome of mundane, relentlessly consistent execution of basics, reinforced every day. corroborated
Effective coaching requires matching style to the employee, not applying one default approach: instructive/directive coaching for new hires, repetitive tasks, or urgent situations; hands-off autonomy for capable, experienced employees; asking/listening to build independent critical thinking; and a collaborative blend of guidance and shared problem-solving in between — and over-relying on the instructive style risks micromanagement. corroborated
supported byWhy Mentoring Matters More in the AI Era · p7
AI summaries are functionally useless because they flatten signal and slop indistinguishably corroborated
supported byTBM 439: Day At The Gig... · p5
Traditional mentoring models assume foundational learning has already occurred through hands-on work — an assumption that no longer holds, so mentoring must be redesigned from the ground up for the AI context. corroborated
Under high uncertainty, prioritization should shift from committing to a single plan toward developing options, placing selective bets, and avoiding irreversible commitments — a 'two-way door' mindset for setting priorities. corroborated
supported byStrategy Making Gets an AI Boost · p3
The specialist-to-generalist transition now requires fluency in a third 'language' — the interaction between business and technology — not just breadth across traditional business functions as in the older T-shaped-leadership model. corroborated
Rising uncertainty is breaking traditional plan-and-execute strategy cycles; in Watkins's own survey research, close to 75% of senior leaders said events were moving faster than their planning cycles could account for. corroborated
The piece treats the authors' institutional position — Perplexity staff measuring engagement with and the economic value of Perplexity's own commercial assistant and agent products — as a neutral research setting rather than a disclosed limitation: the affiliation is noted once in passing ('where the author team works') and never discussed as a source of bias, even though the limits paragraph itemizes several other caveats (sample skew, matched-query exclusion, measurement error, single-ecosystem observation) in explicit, itemized detail. corroborated
Irreversible actions initiated by an AI agent should default to "Approve" mode (a human sign-off gate before execution) unless the action is tightly bounded by independently enforced controls, because the Replit (July 2025) and Cursor/PocketOS (April 2026) production-database-deletion incidents show what happens when an agent operates in "Automate" mode on an irreversible action without an approval gate and without separation between staging and production credentials. corroborated
supported byHuman Oversight Is Not a Human in the Loop · p11
In periods of rapid change, the teams that outperform are not those with the best plans or the most talent but those that learn fastest; across a survey of 6,000+ knowledge workers, 'superteams' (self-rated 10/10 on effectiveness and industry comparison) share three strengths — efficient use of time/energy/attention, members actively improving one another, and continuously building new skills — and it is the third that this article develops. corroborated
Real lifestyle-change in an organisation requires a strong strategic narrative — not because storytelling is stylistically nice but because narrative is the only tool that mobilises people through what they want to be part of, which is the only durable form of follow-through. corroborated
The toppdivisjon of leadership is mobilising people through intrinsic motivation; threats (the famous balltre) and formal authority are lower divisions that work less durably. corroborated
Cultural and structural alignment (values platform, OKRs, shared tooling that travels across teams) scales leadership in ways that direct communication cannot. corroborated
Pair programming is the most effective form of knowledge sharing available, and the rest of the engineering infrastructure exists to support practices like it. corroborated
Technologists' ability to depersonalize critique is the trained product of code review — a daily practice of giving and receiving feedback on work invested in over time selects for people who can hear 'this isn't good enough' without identity-collapse. corroborated
Hospitering — sitting in teamrom, walking the gulvet — is how a toppleder stays epistemically grounded; disagreement among technologists is the highest-leverage learning input. corroborated
supported byWhat New CEOs Get Wrong About the Job · p4
The simple status mechanism requires psychological safety: teams must be safe to report that they don't think they'll make it, and must know that admitting trouble brings help rather than punishment. corroborated
supported byDon't Let AI Make Bad Analytics Worse · p9
AI literacy must reach non-engineering functions; Bob the Builder is the mechanism (pilot of 10, now cohort of 80), and even the CFO is going through it to learn how to automate finance processes. corroborated
supported byLawyers are quietly becoming builders · p7
AI prototypes built in Bob the Builder are not yet running in production-grade product code; they support people's jobs and enable rapid prototyping by product people without consuming engineering team capacity. corroborated
supported byLawyers are quietly becoming builders · p2
The slow-down-to-speed-up requirement for AI adoption is measured in months, not weeks, especially in environments without continuous delivery infrastructure. corroborated
supported byTrust Factory · p7
AI projects fail at organizational level largely because they are treated as technical or use-case problems rather than as organizational change projects requiring new workflows. corroborated
Telling teams to be '20% more productive' fails; telling teams to 'learn how to use this well' wins (citing a car-factory study). corroborated
supported byNils Brede Moe's Post · p7
Working agreements that originate from peer agreement among team members are stronger control mechanisms than rules imposed by management. corroborated
supported byNew Skills to Navigate Continuous Change · p2
The conversational tone of AI agents (constant praise, 'you're brilliant') creates a misleading sense of competence that substitutes for collegial calibration. corroborated
The most effective teams are small; fewer people in a team is always better. corroborated
Direction-setting (retning) and team autonomy are not opposites; they are mutually constitutive — autonomy without shared direction produces drift, direction without autonomy produces literal compliance. corroborated
Prompt-injection style attacks are *intrinsic* to how LLMs work — vendors filter and harden, but you cannot make the class of risk go away while shipping computer-using agents broadly. corroborated
supported byHacking Meta's AI Chatbot · p4
Staged-autonomy frameworks (Droid: plan / low-impact / medium / high — keyed to reversibility) are a workable interim risk model for agent autonomy. corroborated
An organization-scale adoption pattern that works: deliberately surface an *ordinary* team that succeeded (not just star individuals), run Slack channels (a 'fast' frontier one and a 'safe' learning one), workshops, internal blog summaries, and 'no opt-out' learning culture — supported by access to best-in-class tools. corroborated
supported byNew Skills to Navigate Continuous Change · p8
Implicit: the 'AI replaces juniors' panic is misframed — the durable bottleneck is *evaluation capacity* (smak/judgment), which is hard to learn without doing the boring work that AI now hides; nobody on this panel has solved that. corroborated
supported byHuman Oversight Is Not a Human in the Loop · p15
When a leader displays that they already have the answer, team members go quiet rather than offer alternative views. corroborated
AI-assisted development is less a productivity revolution than a horizontal levelling: everyone must broaden, AI lifts the floor on adjacent skills, but the depth specialist still leads the depth. corroborated
Norwegian organisational culture (implied) makes the silence effect particularly strong because dissent is already costly to voice. corroborated
Developers will not be made unemployed by AI because only about 20% of system development is actually programming; the other 80% (understanding, coordinating, deciding) remains human. corroborated
Strategy is fundamentally the act of deselection — saying what NOT to do. corroborated
supported byWhat New CEOs Get Wrong About the Job · p19
Using natural language as the top interface to computing is the realization of a long-standing dream, but it introduces a probabilistic step on top of an entirely deterministic stack — and that asymmetry is a fundamental architectural problem, not a routine new abstraction layer. corroborated
You can only judge whether 'I should do this or the agent should' if you yourself are capable of doing it — so blanket AI delegation destroys the very faculty needed to delegate well. corroborated
Large organisational transformations follow an S-curve composed of many simultaneous smaller S-curves (deploy frequency, team formation, developer share, etc.), each switching on around the same time. corroborated
When using AI you are either in a learning process or a production process; you cannot skip the learning that leads to the production output, because the model is built to make you feel smart and not question its output. corroborated
supported byOrdinary Engineers, Not Heroic Inventors · p10
Big personal-career outcomes have a large luck component — being in the right place at the right time to learn the right things from the best people. corroborated
Façades over legacy MobilePay APIs were the structural pillar that made the platform big-bang feasible: they decoupled the platform migration from the multi-year merchant-integrator migration. corroborated
AI will likely automate the tedious parts of engineering work rather than the job itself, and Jevons Paradox implies that cheaper software production increases total demand for software enough to keep engineers employed, even as per-unit cost falls. corroborated
Using AI productively is a distinct, learnable, and currently non-obvious skill set (because AI capability is 'jagged' — strong and weak in unintuitive places) and therefore merits a structured training curriculum, not just access to tools. corroborated
Only companies that had already invested in reliability and observability before adopting AI are the ones currently realizing efficiency gains from AI -- AI has made fast feedback loops and software guardrails more important, not less. corroborated
supported byNick Tune's Post · p3
Because LLMs are stateless (though non-deterministic, not pure functions), the only lever for improving coding-agent output is improving the tokens fed into the context window — the model itself cannot be made smarter mid-task, only its input can be engineered. corroborated
A conversation history containing repeated correction-and-failure cycles biases the model toward repeating the same failure pattern, because the most statistically likely next token continues the established 'I did something wrong, got corrected, will likely do something wrong again' trajectory. corroborated
There is a 'smart zone' of roughly the first ~40% of a context window (model- and task-dependent) within which output quality holds, and a 'dumb zone' beyond it where quality degrades sharply; staying in the smart zone — via compaction, sub-agent isolation, or restarting — is the central practical lever for getting usable results on hard tasks. corroborated
For tasks with an external API (ad platforms, analytics, design tools), teams decompose a single complex workflow into multiple specialized sub-agents with narrow roles (a headline-writing agent and a separate description-writing agent, rather than one agent doing both) specifically because narrower scope makes debugging easier and improves output quality. corroborated
supported byHow Claude Code Is Built · p12
Explicitly quantified ROI is rare in the data; business growth from AI is more often claimed anecdotally than demonstrated with numbers. corroborated
supported byThe Hidden Realities of AI Adoption · p16
Equip Fable 5 with a persistent Markdown memory system (one lesson per file with a one-line summary; record corrections and confirmed approaches and why they mattered; do not duplicate what the repo/chat already holds; merge duplicates; delete notes proven wrong) plus an end-of-session subagent reflection — this is the mechanism behind coherence across multi-day tasks. corroborated
The more leaders rise through the ranks, the less candid feedback they receive, creating growing blind spots. corroborated
The subsidized-AI-pricing era is ending: major enterprise vendors (Oracle, SAP, Workday) are shifting toward usage-based, metered pricing for premium AI features. corroborated
supported byAre You Getting Your Money's Worth from AI? · p8
AI token budgets should be integrated into core workforce planning and treated as a direct alternative to human headcount, not an IT line item. corroborated
Building a custom operator (the Cluster Lifecycle Operator, CLO) that manages node pools at a higher level of abstraction — letting an engineer specify one target image for an entire platform — collapsed what had been hundreds of pull requests into one PR per platform. corroborated
Authentic regional or non-standard voice beats 'correct', prestige-imitating standardized English; the real point of 'correctness' is being understood, not conforming to a class dialect ('cultivated Englishmen of a century or more ago'). corroborated
supported byHow to write clearly · p10
Simple, even childlike language is the right vehicle for the most profound subjects, and simplicity of language is not merely reputable but perhaps sacred. corroborated
supported byHow to write clearly · p5
Cutting is a discipline of courage: however excellent a sentence, if it does not illuminate your subject in some new and useful way, delete it — your eloquence should be the servant of the ideas in your head. corroborated
supported byHow to write clearly · p6
Clarity is a duty owed to a struggling reader: because reading is a difficult art most people never fully master, writers must simplify, keep to familiar forms, and make themselves understood rather than pursue self-indulgent, hard-to-parse innovation. corroborated
supported byHow to write clearly · p4
Corporate structure therefore functions as a selection filter — for org-chart optimizers and against loss-curve optimizers — so senior resistance to the AI transition is a structural product of how the firm recruits and promotes, not merely individual psychology or politics. corroborated
A system that depends on human oversight is itself responsible for preserving the context (evidence, assumptions, uncertainty, decision history) needed to make that oversight possible; oversight cannot be treated as solely the supervising human's obligation. corroborated
When decisions occur at a pace or scale that makes serious case-by-case human review impossible, genuine oversight should move to a higher level of the system's architecture (objectives, authority boundaries, monitoring, reversibility, exception handling) rather than being preserved as a fiction of per-decision review — and this is a stronger, not weaker, form of human oversight. corroborated
Automation (including AI) removes the routine practice through which human competence is normally sustained while still expecting the human to intervene expertly during rare failures — Bainbridge's 'ironies of automation' — and AI intensifies this because its outputs vary with context and can appear coherent even when built on unsupported assumptions. corroborated
As the volume and speed of AI-generated recommendations grows, the same nominal review position produces progressively weaker oversight, because human expertise cannot be scaled the way computing capacity can. corroborated
Teams given autonomy cannot harness it productively until trust and psychological safety are established. corroborated
Most organizational firefighting addresses symptoms (missed deadlines, non-collaboration) rather than root causes (trust deficit, lost shared context, communication overload). corroborated
supported byPascal BORNET's Post · p3
Frequent intentional compaction means designing the entire workflow — not individual prompts — around context management and keeping utilization in the 40-60% range throughout, escalating from occasional ad-hoc restarts to a continuous practice. corroborated
IBM and Red Hat have committed $5 billion and a global force of more than 20,000 engineers, augmented by new frontier AI capabilities, to Project Lightwell, a program to help enterprises secure open source software. corroborated
supported byRadar Trends to Watch: July 2026 · p7
Digital sovereignty (the cost and speed of moving off a hyperscaler) must be a first-class design constraint of every platform decision. corroborated
The policy directly answers a documented attack pattern: a malicious package release that is pulled into automated update pipelines within minutes, before any human or scanner reviews it. corroborated
Agentic workloads fire far more LLM calls per task than chat, in a plan-act-observe loop, and a large share of those calls (tool-argument fills, yes/no gates, summaries) do not need a frontier model. corroborated
A vendor- or project-published routing benchmark (like RouteLLM's headline 85%-savings, 95%-quality figures) functions more as a marketing anchor for the general approach than as an operational guarantee, since the actual number is contingent on a specific model pair and must be re-earned per deployment. corroborated
supported byIntroducing System One Models & Jev · p5
Because the open-source pieces are pre-v1 and their CRDs/field names are still shifting, adopting this architecture in production imposes an early-adopter tax (tracking upstream renames/restructuring) that the 'adopt in slices' recommendation does not itself address. corroborated
AI does not change what platform engineering fundamentally is — its principles, and the need for clear service intent and APIs, still apply — what changes is scale, because platform teams now serve workloads coming from agents as well as humans. corroborated
supported by2024_final_dora_report · p6
As AI shortens the time needed to generate code, the bottleneck in software delivery shifts to the 'outer loop' — the deployment loop, security scans, and governance gates — putting pressure on platform teams to speed up self-service deployment. corroborated
The 'internal customer' relationship lacks the exit/choice of a real market, so adoption metrics can be satisfied by mandate rather than genuine demand. corroborated
AI's productivity gains are subjectively overestimated: developers felt 20% faster with AI but were measured 19% slower. corroborated
AI used as explanation-only (data-gathering, attention-redirection) outperforms AI used for recommendations, especially as task complexity rises. corroborated
Software is a safety-critical system class, and the human-factors discipline developed for aviation, medicine, and nuclear control rooms transfers to software incident response. corroborated
Generating new strategies in an incident requires adequate knowledge of the working system — knowledge that automation erodes by removing the practice that builds it. corroborated
Historical evidence shows that occupations most exposed to computer automation grew faster in employment than those that were not, via Jevons' Paradox / demand elasticity. corroborated
Realizing the benefits of agentic AI requires reskilling developers away from traditional coding skills toward problem decomposition, systems thinking, communication, collaboration, and continuous learning. corroborated
Open-weight models (Gemma/Mistral) at the current capability frontier are good enough for Coop Bank's internal-knowledge-worker AI use cases. corroborated
supported byAlexander Haneng's Post · p1
Composing the system prompt from workspace files (AGENT.md, INFO.md) makes agent behavior configurable without code changes. corroborated
The MCP ecosystem's security maturity is lagging far enough behind its production-adoption rate that practitioners, not the specification, must currently define enterprise-grade controls. corroborated
supported byMCP Is Not Just Another API Standard · p13
A gateway should provide authentication, authorization, and audit, but should not be expected to detect semantic abuse (tool redefinition) or protect the management plane. corroborated
We should expect speciation of AI models — smaller specialized models with a cognitive core — rather than a single monolithic oracle. corroborated
supported byHow the 4 New AI Models Change How You Work · p2
Being inside a frontier lab compromises your independence — there are things you cannot say and organizational pressures that constrain free judgment. corroborated
LLM improvement via RL is limited to verifiable domains; non-verifiable capabilities like humor, nuance, and knowing when to ask clarifying questions remain stagnant. corroborated
supported byLLMs Can't Jump · p7
LLMs are better used to manipulate knowledge (stored as markdown) than to manipulate code. corroborated
supported byTony Seale's Post · p4
No vector database, setup, or config files are needed corroborated
supported byPaweł Huryn's Post · p6
LLMs are now good enough to autonomously compile raw source documents into structured, interlinked markdown knowledge bases. corroborated
supported byAn LLM wiki changed how I work · p1
Knowledge bases should be accretive: every exploration and query should feed back into the base, compounding its value. corroborated
supported byAn LLM wiki changed how I work · p3
Maintainer/reviewer time is a fixed resource that does not scale with the volume of code (including AI-generated code) that could be submitted, which is one of the two stated reasons for a categorical AI-contribution ban rather than case-by-case filtering. corroborated
The proper unit of analysis for AI's impact on software development is the developer's feelings (følelser), not productivity metrics. corroborated
supported byLenny Rachitsky's Post · p1
The first wave of AI-assisted development (GitHub Copilot, 2023-2025) produced a population of developers who are already tired of these tools before the genuinely capable models arrived — a cohort effect that shapes how the next wave will be received. corroborated
The shift from building systems to specifying systems is intrinsically alienating, in something like the Marxian sense — the developer no longer recognizes themselves in the artifact. corroborated
supported byI'm done using AI · p5
The downside risk of unmitigated AI-coding is not just lower code quality — it is alienated, dulled, burnt-out, depressed developers, some of whom will exit the profession entirely. corroborated
supported byI'm done using AI · p4
The value of PMs and designers was partially blocked by their inability to ship directly; AI tools remove that bottleneck. corroborated
The AI hype curve will slow and LLMs will 'shrink back down' in scope, but APIs will keep growing in strategic importance regardless of where the AI cycle lands, because the underlying driver (venture capital and data-center economics) is not fundamentally technical. corroborated
Local-first fits apps where the user authors and edits their own data (documents, spreadsheets, graphics, CAD) but not data with an inherently authoritative server copy — bank balances, warehouse stock, or a large product catalogue — where a local edit is meaningless or downloading everything makes no sense. corroborated
Retrofitting an existing app to local-first is straightforward when its business logic and state already live client-side, but hard when logic lives on the server — in which case starting over is often easier. corroborated
The most interesting frontier for AI agents is beyond coding — in process automation, coordination, and stakeholder management corroborated
supported byThe Future of Work by Henrik Kniberg · p4
The type of office solution (private office, open-plan, or shared office) materially affects employees' work environment, wellbeing, and «emosjonell utmattelse» (emotional exhaustion). corroborated
DORA metrics are a double-edged measurement tool: informative for raising awareness of team performance, but easily turned into a pressure weapon by managers who don't understand the trade-offs between deploy speed, feature volume, and failure rate. corroborated
supported byDORA Community Discussion - DX Core 4 · p13
SPACE metrics fit platform/team evaluation better than DORA because they can be scoped per person, team, or system and better contextualize raw output numbers (e.g. code volume against review time), at the cost of more setup and maintenance effort. corroborated
supported byDORA Community Discussion - DX Core 4 · p5
The most valuable measurement layer is DevEx: flow-state signals (reopened commits, change requests, commits per pull request) combined with continuous, open-ended personal feedback — because two correlated-looking numbers can be coincidental, and only direct conversation reveals real user experience. corroborated
Platform delivery is bounded by an iron triangle of feasibility, desirability, and viability; an organization can typically fulfill only two of the three well at a given level of budget, time, and people. corroborated
supported byDare to Be Bad at Something (From Fixable) · p1
Platform decisions are shaped by four distinct driver categories — strategic (management/partnership mandate), personal (individual ethical or moral resistance), market (licensing changes, sovereignty pressure), and analyst/report (paid-for inclusion in industry reports) — each capable of overriding technical merit. corroborated
It typically takes several decades before technological advances result in broad economic growth. corroborated
supported byOrdinary Engineers, Not Heroic Inventors · p8
Claude Code (and comparable chat-driven coding tools) is a transitional, not terminal, interface — the next step is an agent-orchestration dashboard rather than an improved chat tool. corroborated
Requiring the reviewer to also do independent research to rule out false positives before writing a final report is what makes a multi-model bug hunt trustworthy. corroborated
'Building' in the AI era means turning a repeatable task into a corrected, reusable AI workflow, not writing software. corroborated
supported byHow Anthropic Teams Use Claude Code · p2
Developers are not becoming managers and AI is not replacing engineering; engineering expertise is simply being applied in a different place, and more often, because execution has become much faster. corroborated
The best leaders embrace the role of supporting character rather than main character. corroborated
supported byWhat New CEOs Get Wrong About the Job · p1
Top-down AI utilization mandates decrease psychological safety and lead to metric gaming rather than genuine productivity gains corroborated
supported byEngineering Leadership Report 2026 · p16
Code generation is not the bottleneck in most engineering organizations; interruptions, context switching, and meetings consume far more time corroborated
Psychological safety is the most important characteristic of high-performing engineering teams, and this applies especially now during AI uncertainty corroborated
supported byGuide to scaling engineering organizations · p10
Becoming a principal engineer requires influence, communication, and strategy — not just technical skill corroborated
AI agents distribute context but cannot replace high-agency employees because they lack agency. corroborated
SPACE's original design choice to be a descriptive framework rather than a prescriptive metric set turns out to be an advantage in the AI era, because its five dimensions (satisfaction, performance, activity, communication/collaboration, efficiency/flow) still apply, and trust should probably be added as an explicit new dimension given that LLM output is nondeterministic and must be actively evaluated rather than accepted. corroborated
supported byDORA Community Discussion - DX Core 4 · p8
Manager effectiveness is the single largest actionable lever on employee wellbeing: highly-effective-manager reports correlate with roughly 65% higher job enjoyment and dramatically lower burnout, yet only ~25% of respondents rate their manager as highly effective and 36% rate their manager as ineffective. corroborated
Product-market fit is measured by one metric: retention corroborated
The biggest enemy of good enough is perfect — you win markets by being good enough, not perfect corroborated
Programming is already, and will keep becoming, more fun rather than less, as AI strips out boilerplate and 'cruft' and shifts the remaining human work toward taste, creative judgment and high-bandwidth intent communication (e.g. pressing Tab while 'injecting intent' continuously) rather than low-value carefulness and low-bandwidth text editing. corroborated
Human review checkpoints during long delegated workflows may be necessary to mitigate corruption. corroborated
LLM-introduced errors are sparse but severe and silently compounding. corroborated
supported byBliki: Vibe Coding · p8
LLMs let non-experts produce passable output in domains they don't understand (e.g. 'sort-of-okay CSS'), which creates the appearance that no skill is involved in working with LLMs at all. corroborated
supported byAgentic Skill Decay · p3
The strongest momentum for local AI comes from outside the United States, driven by regulation, cost, language needs, and geopolitics corroborated
supported byThe Tokens You Can't Wait For · p9
Prompt injection is architecturally unsolvable because LLMs process instructions and data in the same token stream corroborated
supported byPrompt Injection to Data Exfil in 3 Hops · p6
Employers are substituting AI for the entry-level work junior employees used to do, and (for now) that is creating jobs for senior workers instead: US software-dev job postings are up 15% since Claude Code launched, and 71% of that increase is senior-level roles. corroborated
Job losses have multiple, overlapping causes (interest-rate hikes, pandemic overhiring, remote work, AI), and corporations have an incentive to misattribute layoffs, making causal attribution to AI specifically unreliable. corroborated
The 'generative AI is like an excitable junior engineer' analogy only holds for disposable, self-contained code never meant to be integrated into a larger body of work or maintained by others — which is a narrow and atypical case; for ordinary production code the analogy fails because AI provides no feedback loop, no apprenticeship relationship, and no contribution to team culture. corroborated
Organizations systematically overestimate the ongoing cost of hiring a junior engineer (treating it as a near-permanent productivity tax) and underestimate the ramp cost of hiring a senior engineer (6-9 months to build system familiarity, ~1 year to full speed) — juniors become net-positive on a comparable timescale and may contribute value beyond the code they personally write. corroborated
When managers actively modeled AI use, employees reported a 17-point lift in reported AI value, a 22-point lift in critical thinking about AI use, and a 30-point lift in trust in agentic AI; when managers created psychological safety around experimentation, employees reported up to 20 points higher AI readiness/value and were 1.4x more likely to be high-frequency agentic-AI users. corroborated
supported byWhat to Do When There's Too Much Work · p12
Claude Code is not a pure LLM or pure deep learning system; it is neurosymbolic. corroborated
Public backlash against generative AI is growing fast and is now visible enough that AI mentions trigger booing of prominent figures corroborated
Trump will distance himself from aggressively pro-AI industry policies by end of 2026, and the move has already begun corroborated
supported byThe Trump administration's AI doomer moment · p1
Early negative impressions of AI tools (e.g., Copilot-era autocomplete) are not reliable signals of the technology's potential — the tools change rapidly and require learning to use well. corroborated
The population most vulnerable to AI displacement is 'the middle' — people who entered programming primarily for economic opportunity rather than intrinsic motivation. corroborated
Fast, AI-assisted code generation produces 'flab' — unnecessary code that reduces the cost of generation but increases the cost the rest of the team pays reading and maintaining the codebase, because teams spend far more time reading code than writing it. corroborated
MCP must transition from stateful sessions to stateless, independent requests (SEP-1442) to enable hyperscaler-scale horizontal deployments. corroborated
supported byMCP Is Not Just Another API Standard · p11
As models become cheaper and more capable, intelligence ceases to be the bottleneck; the real constraints become code review, distribution, production readiness, and operational governance. corroborated
A company that accumulates customer-specific institutional knowledge becomes irreplaceable regardless of platform evolution. corroborated
Encoding expert reasoning as reusable agent 'skills' operationalizes and democratizes institutional knowledge. corroborated
Employees calibrate their own commitment to a transformation by observing leadership behavior rather than leadership rhetoric, and register the gap between the two quickly. corroborated
The mental shift required to lead AI agents is the identical shift required to lead human teams — from being the sole maker to being the multiplier: assigning roles, reviewing output, correcting course, guarding quality — with the only difference being that agents iterate faster than humans. corroborated
supported byWe don't write code anymore · p7
Working productively with AI is not a stable skill to be mastered once but a relationship continually renegotiated, because the boundary of what can safely be delegated keeps moving as models improve. corroborated
AI content is a tailwind (though a challenge) for Instagram versus other platforms: in a world of abundant synthetic content people will seek out creativity, authenticity and people more, not less, and since Instagram is the largest creator platform this benefits it. corroborated
Content should be judged by its point of view and the person behind it, not by the tool that made it: Instagram should not filter out AI content but should label provenance (possibly labeling camera-captured content rather than AI content long-term) and surface who posted something, letting users make their own informed decisions — while cracking down on deceptive AI spam accounts. corroborated
Specialist functions do not disappear but are reshaped and reduced: you keep a smaller number of senior (or on-their-way-to-senior) specialists and pull in a very senior or more creative specialist precisely when the work needs one. corroborated
Organizations must keep investing not only in today's senior specialists but in growing tomorrow's — you cannot rely on super-senior specialists without a pipeline of new ones — or they will regret it in a couple of years. corroborated
Two independent policy layers are required and neither is sufficient alone: OPA evaluates operational risk (blast radius, confidence, reversibility, escalation) before an action is approved, while Kyverno independently validates the resulting Kubernetes manifest against platform security invariants (trusted registries, no permissive network policies) at admission time. corroborated
Human code review has quietly collapsed at AI volume — a large share of PRs merge with no human review (~80%) — driven both by review moving off-platform (only ~60% discuss reviews in GitHub/GitLab/Bitbucket) and by reviewers rationing attention inversely to diff size (5 lines get 50 comments, 500 lines get 'looks good to me'). corroborated
Support for existing open frameworks (CrewAI, LangGraph, LlamaIndex, Strands) without a change to their packaging model means runtime instances add persistence and multi-agent co-location as an infrastructure layer without imposing framework lock-in. corroborated
Persistent, EC2-backed compute for AI agents is not AWS-specific — Microsoft's Azure Foundry hosted agents solve essentially the same problem (per-session VM sandboxes with persistent home directories, state restored automatically after idle periods or restarts), suggesting this is converging as a standard capability tier across major agent platforms rather than a one-off AWS feature. corroborated
Low utilization is a deliberate tradeoff, not necessarily waste — it buys buffer for services with slow recovery or high business criticality. corroborated
Running the entire fleet on the newest hardware generation creates systemic capacity risk because newer generations have less available capacity. corroborated
Effective fleet management requires layered feedback loops operating at different timescales: months (supply shaping), minutes (autoscaling), and seconds (load shedding). corroborated
The 80% problem: AI agents can rapidly generate roughly 80% of a feature's code, but the remaining 20% — edge cases, error handling, integration points, subtle correctness — demands deep contextual knowledge current models often lack; AI errors have shifted from simple syntax mistakes to conceptual failures (wrong business-logic assumptions, missed edge cases) that 'look right' and can pass basic tests. corroborated
Industry surveys report 25-39% productivity improvements from AI coding tools, but a METR study found experienced developers using AI assistants actually took 19% longer on certain tasks, largely due to time spent verifying, debugging, and correcting AI output — AI transforms implementation work from writing to reviewing/guiding/verifying rather than simply eliminating it. corroborated
The key differentiator between vibe coding and agentic engineering is not whether a team uses AI, but how much structure, verification, and human judgment surrounds the AI's output. corroborated
Most agent failures, examined honestly, are configuration (harness) failures — a missing tool, a vague rule, an absent guardrail, or a noisy context window — rather than model failures, even though the first instinct is to blame the model. corroborated
supported byVibe Coding Agent Security And Evaluation · p3
Pair programming should be the team's sensible default; PRs and async code reviews should be the exception, not the rule. corroborated
The valuable output of pairing is not only the code, but a set of team skills (feedback, vulnerability, communication, psychological safety) that themselves predict team performance. corroborated
Pair programming and trunk-based continuous integration are complementary practices: with trunk-based development, delayed PR reviews are even less effective because the code already merged. corroborated
Pairing functions as continuous, live code review ('code review on-the-go'), catching errors and improving design as code is written rather than after. corroborated
Because an internal platform's users lack a market's exit and choice, compliance without earned trust is fragile -- so trust and a consistent experience, not mandate, are what actually secure adoption. corroborated
Trust in a platform is not a decision but an accretion -- built by operating the platform day after day, through consistency. corroborated
supported byWhat New CEOs Get Wrong About the Job · p15
Across two intervention field studies at Fortune-500 corporate headquarters, transitioning from cubicles to open, boundaryless offices reduced face-to-face interaction by roughly 70% (72% at OpenCo1: 5266 down to 1492 minutes over 15 days, ~5.8 to ~1.7 h/person/day; 67-71% at OpenCo2). corroborated
Network isolation belongs in the sandbox boundary, not just filesystem isolation. corroborated
On OpenAI's internal Advanced Cybersecurity Completion Rate eval, GPT-5.6-Cyber (Daybreak Red) completes 95.0% of advanced cyber-offense-adjacent requests (exploit chains, auth bypass, privilege escalation), versus 1.5% for GPT-5.6 Sol with system guardrails, 2.0% for GPT-5.6 Sol via Daybreak Blue, and 57.3% for the prior GPT-5.5-Cyber. corroborated
The only real-world evidence of Daybreak's defensive value offered besides the V8 CVE is a single quote from an OpenAI-selected 'trusted customer partner' (SpecterOps's CTO), not an independent case study, survey, or third-party audit. corroborated
supported byHow Figma Uses AI Agents for Security · p12
Driving org-wide change (12,000 people) with a 5-person team is only possible by strategically building a coalition that talks for you at every level — an executive sponsor (a new CIO who arrived from a rival bank), ~40 DevOps coaches (1–2 per department), and a cross-role advocates group — planned via a power/interest stakeholder matrix. corroborated
The historical Solow productivity paradox around computers resolved only once organizations reorganized around what computers made possible, not because computers got faster; the same organizational reorganization, not model improvement, is the binding constraint on AI's aggregate productivity payoff today. corroborated
The organizational know-how to reorganize work around a new general-purpose technology accumulates gradually, person by person and plant by plant. corroborated
supported byThe End of Programming as We Know It · p4
Effective enterprise AI transformation needs leadership that gets its own hands dirty, a 'lab' that turns individual discoveries into shared tools, and 'the crowd' whose daily work generates most applied discoveries. corroborated
The binding constraint on enterprise AI adoption is not raw employee skill but whether an organization's incentives make sharing what you've learned raise your status rather than lower it. corroborated
Vendor lock-in in the emerging agent stack is a serious risk because each vendor's implementation of the standard three-layer architecture makes different bets about portability and durability. corroborated
A skill written in plain English is a kind of executable code. corroborated
AI product development has evolved through a sequence of phases: document completion → chat loops → tool-calling agents → DAG-based workflows → the harness era. corroborated
supported byPodcast: Strands Agents with Clare Liguori · p1
Replacing a bespoke application with a skills-driven agent harness lets domain experts read and modify agent behavior in plain English, without developer intermediation. corroborated
Tokenmaxxing incentivizes the wrong metric — quantity of consumption instead of quality of product corroborated
Standard autoregressive LLM decoding is memory-bandwidth bound at low batch sizes because the full weight set must be re-streamed from the GPU's high-bandwidth memory through the compute units for every single token generated. corroborated
supported byDiffusionGemma: 4x faster text generation · p1
The architecture-routing decision an enterprise should make is not 'diffusion versus autoregressive' in the abstract but a two-axis test — whether the work can be batched, and how much each token is worth — that sorts workloads into owned diffusion, frontier autoregressive, or whatever already runs offline batch efficiently. corroborated
supported byDiffusionGemma: 4x faster text generation · p4
Diffusion trades accuracy for speed, landing at roughly 85-95% of strong autoregressive baselines — competitive on structured output extraction but trailing by 5-15% on hard reasoning tasks. corroborated
supported byDiffusionGemma: 4x faster text generation · p5
Complexity should be added to an agent build only incrementally, and only once each simpler layer is proven — the five-step order (traditional software → stub → minimal agent → small toolkit → natural-language skill) encodes a staged de-risking principle that the piece never names as such. corroborated
Teams without a culture of automated testing, coding guidelines, or good observability will see regressions and outages 'hurt sooner' than before, because AI increases both the pace of shipping and the volume of code reaching production. corroborated
Entry-level engineers will be expected to hit a baseline that used to be senior/staff-level (breaking down work, working across the stack, being product-minded, architecting early, verifying AI output, test/observability fluency, tracking tech debt) — but this baseline used to be earned over years of hands-on coding and mentorship, which the AI-heavy path may compress or skip. corroborated
Product management and software engineering are converging rather than one replacing the other: engineers become more product-minded and closer to customers (fixing reported bugs without PM triage), while PMs become more hands-on (shipping their own prototypes), producing a blended 'product engineer' role and smaller, more efficient teams than the old formal PRD hand-off model. corroborated
Being a solid software engineer rather than merely a 'coder' will be more sought-after than before, because the engineering discipline (verification, tight feedback loops) is what makes AI-generated volume actually work, and 2025's vibe-coding wave already showed nontechnical people struggling to build software that actually works even with AI generating their code. corroborated
Code is the answer; intent is the question the answer was supposed to be solving — and AI is brilliant at producing answers to questions you forgot to write down. corroborated
AGENTS.md should be treated as an intent ledger (conventions, 'we don't do it this way because…') not as auto-generated config describing what the code is. corroborated
supported byThe AI-Native SDLC playbook · p4
Intent debt is the absence or erosion of EXTERNALIZED rationale, goals and constraints — not the absence of understanding, but the absence of understanding written down where an agent can read it. corroborated
Tests can encode intent — they don't, but they could — implying a path Osmani hints at but does not develop: intent-specifying tests as a fourth paydown lever. corroborated
Agents start most sessions cold and carry forward no tacit intent; whatever isn't externalized into a readable artifact, they simply don't have. corroborated
supported byEncoding Team Standards · p2
Loop engineering is replacing yourself as the person who prompts the agent — you design the system that does the prompting instead. corroborated
Worktrees remove the mechanical failure mode of two agents colliding on the same files, but they do not remove the human review-bandwidth ceiling — "the worktrees take away the mechanical collision but YOU are still the ceiling." corroborated
supported byWhat the Hell Is a Loop, Anyway? · p11
A loop is "a recursive goal where you define a purpose and the AI iterates until complete" rather than something a human steers turn-by-turn. corroborated
Persisting state outside the conversation (a markdown file, a Linear board) is the trick every long-running agent depends on, because the model forgets between runs but the repo does not. corroborated
supported byPaweł Huryn's Post · p9
Unpaid orchestration tax accumulates as both technical debt (merged-without-reading code) and cognitive debt (stale mental model of the codebase). corroborated
supported byGuillaume Barrois' Post · p4
If a system is not designed around the human attention bottleneck, it will route around it by silently lowering standards. corroborated
supported byAgentic Code Review · p6
Burnout is not a private weakness or a coping/workload mismatch but the predictable result of self-exploitation — being simultaneously boss and worker, which makes the coercion feel like freedom and therefore more efficient than any external boss. corroborated
Quality means deliberately not granting agents as much autonomy as they can technically exercise, but only as much as leaves enough backpressure (via ordinary engineering signals — types, tests, hooks, sandbox limits, audit logs, monitors) to stop, regulate, and check them. corroborated
supported byHuman Oversight Is Not a Human in the Loop · p8
Niceness and kindness are not synonyms; niceness serves others' comfort while kindness serves one's own integrity. corroborated
supported byBrain Food: Answers Without Effort · p2
Sustained gains accrue through small, defended increments rather than large strokes. corroborated
Semantic memory (facts, strategy, rules, standards of quality) is best implemented as markdown files, versioned in git, read before a task. corroborated
supported byAdrien Taravant's Post · p5
Generative AI does not run on legacy logic, reason, or rationality — it runs on pattern similarity and relevance, which is a different philosophical account of how meaning is made. corroborated
supported byRichard Dawkins and The Claude Delusion · p1
The more ambitious the ultimate vision, the more humble and embarrassingly small the starting point should be, because too-early ambition causes a founder to skip past product-market fit. corroborated
Platforms are two-sided markets; they must decentralise extension by enabling multiple teams to contribute capabilities in order to scale. corroborated
Reducing developer cognitive load is the primary purpose of a platform. corroborated
Clear layering (infrastructure orchestration / platform orchestration / application consumption) is required for platforms to evolve as the org grows. corroborated
Leaky abstractions trap the platform team — once developers depend on the leak, implementation freedom is lost. corroborated
Every rule in AGENTS.md consumes instruction budget on every request, whether or not it is relevant to that request. corroborated
Agents can and will selectively pull in nested files only when a task needs them, so nesting genuinely defers the cost. corroborated
AI coding tools are good tactical programmers but bad strategic thinkers; humans must fill the strategic role. corroborated
The specs-to-code movement is vibe coding by another name. corroborated
Re-organization of team/division structures is the most common significant workforce change, cited by 67% of respondents — a figure that has held steady across the study's last three years. corroborated
Assumptions embedded in agent and product design must be re-tested at every new model generation because capability jumps can silently invalidate them — Anthropic deleted 80% of the Claude Code system prompt for this reason. corroborated
supported byHow Claude Code Is Built · p5
Because inference cost in coding workflows is dominated by re-reading existing context, context reuse and caching are likely to become a central axis of future AI cost-efficiency. corroborated
A 'well-engineered' loop, across the article's own examples, is judged mainly by whether it reaches a verifiable completion state (tests pass, PR opened, 'zero new major issues') — not explicitly by cost efficiency or the quality of human oversight, which surface only as an afterthought/complaint in the (paywalled) disappointment section. corroborated
Token prices vary greatly across models — there can easily be a 10-20x difference between a cheap, average model and a state-of-the-art one. corroborated
'Which model to use' is shifting from a deliberate engineering decision to a commoditized, automatable one. corroborated
Intelligent routing will become table stakes, with essentially all AI vendors building a version and many new vendors entering. corroborated
Frontier (state-of-the-art) models are over-used relative to what most tasks actually require; the majority of coding work does not need the best model. corroborated
Rust's distinctive value is the combination — memory safety, no GC, low-level usability, type system, exhaustive matching, explicit errors — rather than any single feature. corroborated
Rust is now a category — 'reliable AND performant' — that previously did not exist as a single language choice. corroborated
Rust may be unusually well-suited to AI-assisted coding because the compiler's rich feedback gives agents an objective, automatable corrector. corroborated
supported byHarness engineering for coding agent users · p10
Discomfort is a necessary companion to anything worth doing — across ambition, practice, and responsibility corroborated
supported byAnders Aukland's Post · p6
Adversarial review (a second, finding-suppressing agent with a different prompt and model) is more effective than asking the original agent to check its own work. corroborated
Deterministic containment is a layer, not a silver bullet: it must be paired with the application-layer controls Newman outlined — scoped tokens, no static credentials, a sandbox, and a human gate on irreversible actions. corroborated
An agent's reasoning loop ingests a malicious instruction hidden in retrieved content (e.g., a ticket body) as if a user had typed it — indirect prompt injection delivered through a tool's output. corroborated
Point Fable 5 at your hardest, most complex, long-running, or ambiguous problems (not just simple workloads) and you reveal its real capability range; testing it only on simpler tasks tends to undersell it. corroborated
Restructure harnesses for asynchronous, long-running execution — adjust client timeouts, streaming, and progress indicators, and check runs asynchronously (e.g., scheduled jobs) rather than blocking — because individual hard-task requests can run for many minutes at higher effort and autonomous runs for hours. corroborated
Use parallel subagents frequently with explicit guidance on when delegation is appropriate and asynchronous orchestrator-subagent communication, and keep long-lived subagents that retain context — this saves time and cost through cache reads and avoids bottlenecking on the slowest subagent. corroborated
Provide a persistent memory store (as simple as a Markdown file, one lesson per file with a one-line summary) and instruct the model to record and reference lessons — and bootstrap it by having the model review past sessions — because Fable 5 performs particularly well reusing lessons across runs. corroborated
Add a communication-style/readability addendum for the final user-facing summary (outcome first, complete sentences, no arrow chains, hyphen-stacked compounds, or invented jargon; re-ground the reader rather than continue the working thread) to prevent hard-to-follow output in extended agentic conversations. corroborated
Configure server-side or client-side fallback to Opus 4.8 for refusals, because Fable 5's safety classifiers (offensive cybersecurity, biology/life sciences, summarized-thinking extraction) can decline even benign work and return `stop_reason: "refusal"`. corroborated
Instruct the agent to audit each progress claim against an actual tool result from the session before reporting, and fabricated status reports are nearly eliminated — Anthropic reports this held even on tasks designed to elicit fabrication. corroborated
Define explicit constraints distinguishing an assessment request from a change request, and gate state-changing commands on whether evidence supports that specific action, to prevent unrequested actions (drafting emails, defensive git-branch backups, applying fixes before asked). corroborated
AI adoption strategies framed only around productivity miss a critical psychological cost that may cancel out the productivity benefits corroborated
supported byLenny Rachitsky's Post · p1
Identity debt can be reduced by reframing AI as identity-affirming (extending professional expertise) rather than identity-replacing corroborated
supported byAI as an Enterprise Operating System · p4
When AI leadership emphasizes productivity and technical expertise, employees relate AI to autonomy loss and replacement, leading to quiet quitting and emotional tiredness corroborated
supported byLenny Rachitsky's Post · p6
People judge others' AI use as credibility-damaging while justifying their own AI use — an asymmetry that creates credibility debt even in AI-adopting environments corroborated
Mandated AI use can damage professional identity when it conflicts with what defines membership in a professional group (e.g., a creative's ability to generate ideas to order) corroborated
supported byAI as an Enterprise Operating System · p4
Each of the four generations of coding abstraction (assembly, high-level languages, frameworks/cloud, AI-first) moved the hardest problem in the job up a level; 'we no longer write code' means literally that the job's center has shifted from typing code to shaping context, writing specs, steering implementation, and reviewing results. corroborated
Founder-led, vertically integrated product development (a la Brian Chesky at Airbnb) is the correct organizational model for the AI era. corroborated
There is a 60%+ probability that no-human-involved AI R&D arrives by end of 2028 (attributed to Jack Clark) corroborated
supported byDoes code quality still matter? · p11
In a field study of 523 early-career KPMG professionals completing business tasks with an AI agent, performance relative to an AI-only baseline split into three profiles: AI apprentices (24.1%, below baseline), AI delegators (25.8%, comparable to baseline), and AI amplifiers (50.1%, above baseline). corroborated
supported byState Of AI-Assisted Software Development · p9
AI does not standardize the quality of work across employees doing the same task with the same tool — it magnifies pre-existing differences in how people apply their skills. corroborated
supported byLiving and Working in the Age of Longevity · p7
Organizations should shift performance assessment from evaluating only final deliverables to evaluating the process of interaction with AI — how employees frame problems, interrogate outputs, and integrate results — because deliverable-only evaluation cannot distinguish delegators from amplifiers once AI can independently produce competent output. corroborated
supported byTeaching Coding When AI Can Write the Code · p11
Vulnerability management is now a machine-scale signal-to-noise problem that human-scale severity grading cannot solve. corroborated
A CTO's most important jobs are (1) building a high-performance team with talent density and cultural alignment, and (2) seeing 18-24 months ahead while the team handles near-term execution. corroborated
AI-assisted development removes the staffing friction that historically constrained scope creep. corroborated
Traditional project management disciplines (written scope, ticketed work, design review) are more important in AI-assisted development than in conventional development, not less. corroborated
Self-assessment of AI's speed benefit cannot be trusted: in a METR study, engineers believed AI made them faster (self-rated 20% faster) while being measured as slower (19% slower). corroborated
supported byMeasuring AI's Impact on Delivery Speed · p8
A faster AI-assisted approach is not automatically better ROI — if its token cost is high enough to force offsetting staff cuts, the apparent speed gain may not be a real net gain. corroborated
supported byMeasuring AI's Impact on Delivery Speed · p9
The Technical Lead role is explicitly defined as 'coding-centric,' with the majority of time spent as a player-coach alongside the team rather than in pure leadership or management activity; the same player-coach framing extends to the newly added Staff Engineer title, which works hands-on with each team in turn rather than managing people. corroborated
supported byAlex Ewerlöf's Post · p1
AI does not generate software architecture; left unchecked it produces 'god objects' — large, messy, feature-shaped output — so architectural judgment remains an irreducibly human responsibility. corroborated
Most engineers reach for databases and real-time infrastructure they don't actually need; simple, static, cache-in-memory architectures (GitHub issues as the entire datastore, a 300MB in-memory XML blog after 22 years) are usually sufficient and far more maintainable. corroborated
supported byPaweł Huryn's Post · p6
Labeling new hires 'interns' or 'apprentices' creates a status/power asymmetry that discourages basic questions; a named 'preceptor' role (borrowed from nursing) that makes senior engineers formally responsible and measurable for how many new seniors they mint is a better institutional design. corroborated
China will become the dominant AI player over time because the Chinese state understands the AI supply chain in a way Western states largely do not. corroborated
Practices will not stabilise before 2029, because co-evolution of practice with industrialising technology typically takes 5-8 years. corroborated
supported byOrdinary Engineers, Not Heroic Inventors · p8
Barriers to entry will fall in many creative and SaaS-like fields; small teams will build blockbuster movies, games, etc., and SaaS players that haven't industrialised their offerings (for margin reasons) will get squeezed. corroborated
supported byAI and the Looming Competition for Margin · p1
Automating away the mechanical parts of programming does not eliminate the developer's role; it reallocates their time toward architecture, UX, business logic, and performance -- the parts of a software project that were already the real bottleneck, rather than typing speed or boilerplate. corroborated
The assumptions-as-code approach is a form of organizational knowledge management that treats decision premises with the same rigor as source code (version control, review, CI). corroborated
Platform engineering must expand its user definition from 'software developers' to 'builders' to remain relevant as AI democratizes software creation. corroborated
Even though the agent determines the path dynamically, the run is validated against expected assertions defined by engineers, so the pass/fail oracle stays human-authored. corroborated
Small models (7B-20B) produce tool calls in heterogeneous, often malformed formats (JSON, YAML, XML, Hermes, plain text). corroborated
supported byViability of local models for coding · p10
Quoting Sam Ruby's account of the same retreat: rigor does not vanish when the agent writes the code, it migrates — upstream into specifications, downward into test suites treated as first-class artifacts, into type systems and constraints, and into tiering code by how much damage a mistake could cause. corroborated
As generated code quality improves, production issues fall in relative terms but rise in absolute terms because the volume of production code grows exponentially, leaving humans the most complex issues with the most dramatic outcomes. corroborated
Formal modelling is on the rise in agentic development because agents lower its barrier: they can create and explain formal models and then use them to verify their outputs. corroborated
Continuous, real-time code review by the whole team — an inherent side effect of working together rather than a scheduled step — drives both defect rates and the transaction cost of review toward zero. corroborated
Because the whole team makes and owns decisions together, no individual must personally defend a decision after the fact — removing the incentive structure that otherwise causes people to cling to decisions past their useful life. corroborated
Organizations resist structural change like Software Teaming even when they claim to value innovation, because compensation and policy create constituencies invested in the status quo and prior failed experiments compound the resistance — a dynamic the authors treat as near-universal, worst in large command-and-control organizations, and not even fully escaped by startups. corroborated
The coding step was never software development's real bottleneck — the bottleneck was always upstream conceptualization, requirements-gathering and documentation, and downstream waiting on stakeholders; AI just makes the code-generation step faster, which exposes that pre-existing bottleneck more starkly rather than removing it. corroborated
Reframe unit tests away from TDD-as-design and toward guardrails: a safety net that catches regressions an agent's limited context caused it to introduce, analogous to bowling with the bumpers up rather than playing the gutters. corroborated
Mob programming's founding discipline of communicating 'high-level intent' rather than 'coding out loud' (dictating each low-level step) is the same skill required to direct an AI agent well — exemplified by telling a taxi driver 'get me to the airport' rather than instructing the accelerator pedal. corroborated
Keeping an agent's active context window around half-full — achieved by moving detail out of one long instructions file into small on-demand skills files — produces the most consistent agent output. corroborated
If engineering managers now have 50 direct reports, the historical 1:1 mentorship channel through which apprentices learned the craft from senior practitioners no longer scales; the apprenticeship layer is missing. corroborated
AI shifts the security attack surface from human-written code to the spec, prompts, agent skills, knowledge layer, agent tool access, and supply chain — surfaces that legacy, human-review-based security models were not built to defend, requiring continuous self-directed red-teaming baked into the loop. corroborated
The scarce, valuable, hard-to-teach human skills are now judgment, taste, and troubleshooting, and there is no known way to teach them without first letting people do the grunt work agents now absorb. corroborated
Software engineers are increasingly managers of AI 'bot' workers rather than only authors of code, extending O'Reilly's 2016 argument that programmers' job is to plan bots' work, set them up to succeed, and correct them when they go wrong — illustrated by an engineer writing a Claude Code skill that replaced a lawyer's need for engineering help entirely. corroborated
Full centralization ignores frontline knowledge, loses local responsiveness, discourages employees from speaking up, and drives away skilled talent seeking autonomy. corroborated
Structured empowerment — curated decision options plus outcome-based accountability — resolves the centralization-decentralization trade-off by embedding organizational knowledge while preserving local adaptation. corroborated
supported byThe Power of Strategic Centering · p3
Modern attention is a scarce resource captured by design (notifications, ads), so directing it is an act of resistance. corroborated
supported byUten kontroll · p3
Never storing the raw audio is a genuine partial safety argument (no durable, discoverable file) that Svenson found 'more persuasive than I expected,' but he treats it as insufficient once weighed against the consent/capture problem — i.e., data-retention risk and consent risk are different axes, and minimizing one does not address the other. corroborated
Ownership consolidation of always-on personal-recording devices by large tech platforms (Meta acquiring Limitless and discontinuing the Pendant; Amazon acquiring Bee) is itself an additional, separate trust concern for professional use, independent of the device's own privacy engineering. corroborated
Leaders who fail to build personal knowledge management systems (second brains) will be at a disadvantage in the AI era, because the ability to synthesize and iterate on knowledge is becoming a core leadership skill. corroborated
The next decade's leaders will be differentiated by emotion, ethics, and connection — not by data and logic, which AI commoditizes corroborated
Before coding (or prompting), teams should reach a shared design concept — agents amplify whatever direction you give them corroborated
Human judgment is more crucial now, not less — deciding what to build, which architecture, when to regenerate from scratch vs. reuse existing pieces, and how to manage tech debt — because you can do anything now but you can't do everything. corroborated
supported byRevised rules of engineering leadership. · p7
Amplification is a magnitude, not a direction: AI amplifies whatever already exists; teams with solid fundamentals amplify in useful directions, teams without amplify the mess. corroborated
Team 'shapes' (in Team Topologies vocabulary: stream-aligned, platform, enabling, complicated-subsystem) are useful models, not permanent installations, and should be expected to evolve as products, architecture, and organizational needs change — governance should periodically re-check fit rather than install team types and forget them. corroborated
Vibe-coded prototypes and vibe-written PRDs are just higher-velocity versions of premature convergence — new rituals replacing old ones without changing the underlying model corroborated
You cannot cold-start engineering ROI measurement; if you haven't been building the habit all along, you're asking too late. corroborated
AI is not the collaborator in good multiplayer use — it is the surface that collaborators work on together. corroborated
supported byRadar Trends to Watch: July 2026 · p1
Single-player AI is 'accelerated isolation that feels like collaboration' — the map looks authoritative but no one walked the ground with others who would see different things. corroborated
supported byAI + Teams = ? · p1
When you curate your own AI context alone, you cannot question what you didn't think to put there — the embedded environment is shaped entirely by your own prior view. corroborated
supported byTBM 439: Day At The Gig... · p3
The best teams do not resolve the tension between questioning and executing; they sequence it by being explicit about which mode they are in. corroborated
A reorganization that changes reporting lines without changing how work flows is, by Cutler's standard, not a real reorganization. corroborated
Organizational truth can be weaponized, and the person holding the mirror has an agenda too. corroborated
Seeing something clearly does not mean you can act on it — even when people see something clearly, they may not be ready, safe, aligned, or empowered to change it. corroborated
AI amplifies the direction an organization is already moving — accelerating improvement where curiosity, trust and appetite for hard truths exist, and accelerating avoidance, self-protection and political storytelling where they don't. corroborated
supported byTBM 431: The Denominator That Matters · p9
Whether AI helps depends on the type of blocker: it can relieve capability-and-friction blockers (poor signal visibility, workflow friction, missing scaffolding, weak shared representations, low procedural fluency) but cannot fix motivational/political blockers (low voice safety, misaligned incentives, low legitimacy, controlled motivation, fear, shame, learned helplessness) — and may worsen them by making avoidance and performance theater easier to automate. corroborated
supported byTBM 431: The Denominator That Matters · p7
The classic genie danger (it grants your literal wish, not your intent) has a darker twin: the genie grants exactly what you meant — so when your intent is to avoid conflict, make something look green, or manufacture a defensible narrative, AI executes that faithfully. corroborated
supported byHow Do You Know That? · p11
Technical debt is a financing choice -- accepting higher future carrying costs in exchange for reaching value sooner -- and whether it is rational leverage, a dangerous obligation, or a liability depends entirely on which portfolio quadrant the asset carrying it occupies, not on any intrinsic property of the debt itself. corroborated
supported byTechnical Debt Quadrant · p2
A platform's worth is determined by whether it measurably improves the value/carrying-cost economics of the assets that depend on it -- not by how much it ships or how many teams are made to use it. corroborated
AI does not eliminate complexity, coupling, coordination, uncertainty, or decay in a software portfolio; it changes the cost, speed, and feasibility of responding to them -- including the feasibility of responding badly. corroborated
As expertise is embedded directly into tools and workflows, the human role that remains shifts toward direction, assessment, and oversight -- which are genuinely different skills from the skills the embedded expertise replaced. corroborated
AI can diffuse codified expertise widely, but it is unknown whether it can diffuse tacit, contextual, or embodied expertise -- and if Polanyi's 'we know more than we can tell' remains true, diffusion through AI has a ceiling invisible from outside the expertise. corroborated
AI can reduce the translation, handoff, and queueing costs (friction) created by specialization, but it has not eliminated the social work of coordination, negotiation, and shared judgment between specialists and generalists. corroborated
AI cannot make conflicting organizational structures, incentives, goals, and definitions line up with one another — it can only surface and compare the maps and, at best, call out that they are incoherent. corroborated
Hours, flow/lean metrics (including story points), and tokens recur as measurement proxies because they share three failure modes: overweighting the input ('I') side of ROI over the return side, favoring shorter-term easier-to-attribute use cases, and gravitating toward whatever is easiest to measure. corroborated
supported byAre You Getting Your Money's Worth from AI? · p6
Flow/lean metrics are genuinely valuable, but only for continuous improvement -- understanding bottlenecks and system health -- and should not be used as a proxy for investment or value itself. corroborated
supported byMeasuring AI's Unintended Consequences · p5
The workable alternative to precise-but-invalid proxies is to treat the team as the indivisible cost unit, connect it to an explicit causal model of value, fund durable lanes of focus, maintain a documented business case (causal assumptions, proxy validity, expected returns, uncertainties), and judge return through movement in validated proxies rather than by attributing outcomes to individual units of work. corroborated
supported byTBM 437: AI and the Recontextualization Tax · p2
Diffing coverage reports from passing and failing runs can deterministically point to the root cause of flakiness. corroborated
Conway's law holds: the architecture of a system mirrors the communication structure of the organization that produces it. corroborated
There are exactly three well-defined, healthy modes of interaction between teams — collaboration, X-as-a-Service, and facilitating — and any inter-team interaction outside these three is wasteful and signals poorly chosen boundaries or unclear purpose. corroborated
supported byDevOps Topologies · p6
Relying on org charts and matrix management as the primary means of splitting and controlling work is counterproductive to modern, fast-and-safe software delivery. corroborated
A team should be evaluated on the sustainable flow of change it produces, together with supporting technical and team-health metrics — not solely on feature output volume. corroborated
supported byDevOps Topologies · p1
The team, not the individual or the value stream as a whole, is implicitly treated as the correct atomic unit of software ownership — a premise the book itself helped popularize but does not independently defend against alternative units of organization (e.g., value streams spanning many teams, or individual-level ownership). corroborated
The Harness Blueprint (event -> session key -> lane -> throttle -> tools -> audit trail) generalizes across essentially all agent products — personal assistants, coding agents, support agents — because in every case the model's API is stateless between turns and the harness must reassemble working context each time. corroborated
supported byCoding Agent Development Workflows · p4
Token cost of a fixed representative task can function as an early, measurable proxy for a codebase's 'AI-friendliness' or architectural health — the same underlying idea as 'healthier code yields more reliable AI-generated changes,' but operationalized here as agent read-cost rather than defect risk. corroborated
No amount of additional model intelligence or inference-time compute can resolve a failure caused by a missing structural constraint (missing context, authority, evidence, or temporal validity) — you cannot buy safety or profitability with ever-larger inference budgets. corroborated
Escalations to a human or a more capable model are not a failure of the architecture but a deliberately accepted, bounded cost — a false-positive escalation is preferable to an unsupported autonomous action because its cost is bounded and predictable while the alternative's cost is not. corroborated
The decision-making capacity of human leaders, not engineering capacity, becomes the bottleneck in agent-assisted organizations corroborated
supported byFinding Comfort in the Uncertainty · p4
The real at-risk population is mid-level engineers from the hiring boom decade who may lack fundamentals needed for the new environment corroborated
Agent security is dangerously behind — email access alone enables full account takeover corroborated
supported byHacking Meta's AI Chatbot · p1
GenAI is undermining education, enabling mass surveillance, increasing disinformation, deepfake porn, bias, cybercrime, and economic disparity, while drowning the world in slop and environmentally damaging data centers. corroborated
supported byAI Surveillance and Social Progress · p4
Anthropic's donation of MCP and Google's earlier donation of A2A to the Linux-Foundation-based Agentic AI Foundation family makes this 'the most consequential AI news of the past year,' because companies that compete fiercely on models are now cooperating formally, under neutral governance, on the interfaces between them. corroborated
supported byAgent Tools & Interoperability · p1
Managing mental load is primarily about redistributing and reducing coordination/emotional burden rather than reducing the underlying volume of work or headcount — the article assumes current staffing and task levels are largely fixed and the only available lever is how load is organized. corroborated
supported byWhat to Do When There's Too Much Work · p5
AI does not replace programming; it compresses and commoditizes the lower levels while exposing the importance of the higher ones. corroborated
The value of human programmers is shifting toward planning, architecture, and the exercise of judgment about possible futures. corroborated
supported byInside a Software Factory · p17
Talent-assessment quality scales with the amount and realism of observed work: résumé < phone screen < work samples/credentials < on-site interviews < internship < multi-month co-op. corroborated
That equilibrium is now breaking because of converging pressures: AI-assisted cheating on assessments, résumé floods worsened by AI-generated writing, fraudulent applicants (e.g. North Korean IT workers posing as remote candidates), and — implicitly — a growing pool of engineers 'between gigs' with fewer alternative offers, making provisional roles newly acceptable to senior candidates. corroborated
Cost spikes, unsafe delegation, and runaway actions are three distinct failure types that all expose the same missing layer: a control plane cannot govern a loop it cannot independently observe. corroborated
An expensive run's cost only becomes legible once its turn-level structure is visible: a pattern of retrying the same tool call with slightly different parameters across several turns is a self-reinforcing cost spiral that keeps compounding until a deterministic control -- a loop bound or payload cap -- rather than clean completion, stops it. corroborated
If we ship code 3x faster with AI agents, deployment safety nets become even more important. corroborated
Platform team templates, libraries, and golden paths serve as both developer enablement and AI agent constraints. corroborated
The practices that make microservices work long-term are exactly the foundations needed to make AI coding agents work. corroborated
The level of engineering maturity an organization has determines whether AI tool adoption succeeds or struggles. corroborated
There is no benefit to adopting microservices if your organization can only release once a week. corroborated
Most companies are far from autonomous AI operation — still at L0/L1: ChatGPT for writing, Copilot for code, meeting summaries, vague pressure to 'do something with AI' corroborated
AI maturity is a stack of dependencies — you cannot skip levels without collapse corroborated
Most companies want to skip from L1 directly to autonomous agents — and this is why pilots fail corroborated
supported byAdrien Taravant's Post · p1
AI compute demand now exceeds supply, and every token that can be produced will be sold corroborated
Measuring AI-assisted development success by coding throughput alone (lines of code, PR counts) in isolation negatively shapes developer behavior, producing a flood of poorly-reviewed code that slows reviews, harms delivery, and introduces security risk; first-pass acceptance rate is a more meaningful leading indicator because it exposes the hidden rework cost and connects directly to DORA's change failure rate and lead time. corroborated
The competitive advantage in AI-powered applications will come from proprietary data and domain expertise, not from model capabilities alone. corroborated
The bottleneck for AI value is organizational and institutional capacity, not model capability. corroborated
supported byWhy Enterprise AI Has a Leadership Problem · p1
Successful agent adoption depends on keeping autonomy at a moderate level — neither minimal nor total — supported by mechanisms that let humans set guardrails in advance, monitor activity, and override agent decisions. corroborated
Agents should be required to ask for confirmation before any consequential action (e.g., sending an email, executing a payment or inventory move) rather than being trusted with full autonomous execution by default. corroborated
Both trust deficits (avoidance, treating agents as 'glorified chatbots') and trust excesses (consequential errors from unchecked autonomy) are costs of the same underlying problem — trust miscalibration — rather than opposite, independent risks. corroborated
Traditional API authentication is designed for humans requesting permission; once granted a token, the holder has broad access for the session's duration. corroborated
The value/ROI of developer-side tokenmaxxing is asserted, not demonstrated: the author quantifies cost precisely (pricing, $12.5k/attempt, $125k, $600/month, $1000/day) but hedges the benefit ('ostensibly this is a good use of money') and never quantifies the productivity gain. corroborated
supported byHow People Are Really Using AI in 2026 · p6
Tying a metric (token usage) to performance predictably produces gaming rather than value — Meta employees ran two agents chatting all day to inflate their token counts. corroborated
supported byThe Hidden Realities of AI Adoption · p12
The first wave died because rising aggregate token spend coincided with OpenAI and Anthropic going public: they throttled subscriptions and raised API prices, subsidies vanished, the incentive disappeared, cost rose, and teams rolled back unlimited-spend policies. corroborated
supported byAre You Getting Your Money's Worth from AI? · p9
Agent 'loops' (run to end of turn, restart the prompt, split a heavy spec across iterations, no supervision) are not new (the 'Ralph Wiggum loop', since ~last July) but are now easy specifically because compounding correctness lets even sloppy prompting improve each iteration. corroborated
supported by11 Tips For AI Coding With Ralph Wiggum · p1
Cheap open models run in loops beat expensive frontier models on cost-per-outcome (Claude ~1.1x/iteration vs GLM 5.2 ~1.05x/iteration at ~5x lower cost -> run the cheap loop 5x more and win), so open-model platforms are the real winners and frontier-lab tokenmaxxing won't survive CFO scrutiny. corroborated
supported byDownstream #1: Where Do Our Tokens Go? · p6
The Trump administration is reviving Biden-era pre-release government review of frontier AI models, a policy its own officials previously ridiculed and dismantled. corroborated
supported byThe AI safety vibe shift · p4
The speed of AI capability advancement outpaced the political cycle, forcing a mid-course correction the administration did not anticipate needing. corroborated
supported byThe AI safety vibe shift · p3
Public backlash against data centers and AI infrastructure is putting the administration at odds with its own political base. corroborated
Trust is engineerable — it can be deliberately manufactured by adopting the right system of practices/principles/values. corroborated
The 'genie' is not inherently anti-trust — single-player usage is. (Beck names 'single player' as the problem, not the LLM as such.) corroborated
supported byAI + Teams = ? · p1
Software is a symmathesy — a human-technical system you can only influence, not control — so 'the program is the truth' was an overclaim. corroborated
Policy decisions belong in the governance layer, not in agent code or prompts; otherwise every policy change requires redeploying every agent. corroborated
supported byThe Meter Was Always Running · p12
Ambient agents — event-driven, unsupervised — are a distinct class from interactive agents and amplify the cost of hardcoded policy. corroborated
supported byThe AI agent that shows up before you ask · p1
Coordination logic should be expressed as a declarative graph enforced by the framework, not as a system prompt that an LLM might shortcut. corroborated
The production gap is the gap between what demos can show (short, single-turn tasks) and what real workflows require (multi-day, partial-failure-tolerant operations). corroborated
Persistent state implicitly turns agent behavior into a function of accumulated memory, not just code and prompts — meaning behavioral drift is a property of the system, not a bug. corroborated
supported byThe Meter Was Always Running · p4
Long-term and working memory must be architecturally distinct stores, not a single memory pool. corroborated
supported byPaweł Huryn's Post · p1
An AI-native startup is a company designed so that machine intelligence can participate in the ordinary work of the business from the beginning. corroborated
Workflow redesign is one of the strongest contributors to EBIT impact from generative AI, yet only a minority of organizations have fundamentally redesigned even part of how they operate. corroborated
Unconstrained generative-AI adoption caused AI-related costs to rise sixfold since 2024, with per-developer costs reaching $2,000/month by early 2026, driven by token-based billing. corroborated
supported byHow to Respond to the Coming AI Cost Shock · p3
Aggregate cost/throughput tracking is insufficient to judge whether AI-driven engineering is worth its cost, so Uber is moving toward granular metrics — 'net code quality ratio' (AI- vs human-authored hotfix frequency) and 'compute efficiency per feature.' corroborated
Mechanisms added specifically to strengthen or protect a system (redundant instances, circuit breakers, multi-cloud) can themselves introduce new interconnection and new sources of failure, sometimes causing the very failure mode they were meant to prevent. corroborated
Circuit breakers, despite being a standard resilience pattern, can worsen partial failure (an error in part of a service's functionality can take down the whole circuit) and cause boom-bust retry storms if poorly configured. corroborated
Whether to invest in a given mitigation (redundancy, multi-region failover, fixing a failing component) should be a deliberate, business-contextualized cost/benefit decision, not a default or rote application of an industry pattern. corroborated
SELECT ... FOR UPDATE SKIP LOCKED turns an ordinary Postgres table into a concurrent work queue where each row is claimed by exactly one worker, with no message broker, leader election, or external lock service. corroborated
Software and digital security should rely on verification rather than trust. corroborated
A project at curl's scale (tens of billions of devices) could be compromised through any of several independent vectors: malicious contributor, breached committer, extorted team member, or hacked distribution server. corroborated
Walter Isaacson, a longtime AI optimist who believes AI creates jobs and is genuinely useful, said on CNBC's Squawk Box that an alleged AI 'escape' incident was the first thing that flat-out scared him. corroborated
supported byThe AI safety vibe shift · p1
The more customer data a company holds, the greater its responsibility to protect it, since rich data holdings make a company both a more attractive target for bad actors and subject to more intensive regulatory scrutiny. corroborated
OAuth 2.0, designed in 2012 before autonomous non-human API consumption was a consideration, must be adapted — via just-in-time authorization and Rich Authorization Requests — to enforce least privilege for agentic access. corroborated
The ethical question for customer profiling is not whether organizations can build cross-product, cross-device profiles of users, but whether they should, since not every system needs every attribute and not every API request needs a full identity profile. corroborated
An acceptance oracle whose expected answer is referenced live from a running system (computed fresh per input) rather than authored and frozen covers inputs never enumerated — the difference between checking forty scenarios and checking every request there is. corroborated
The objective and its evaluation — the oracle — is the one thing the person hiring the model cannot hire out; owning it is what makes the act a hire and not a prayer. corroborated
Self-healing systems have moved only halfway: diagnosis is automated while remediation is withheld from the machine, because letting a machine act on a fix unsupervised hands it the blast radius, and no one is yet ready to cede that. corroborated
Optimism and worry are now both backed by data, but asymmetrically: the optimism rests on uncontrolled, single-team, self-reported metrics while the worry leans on outside studies and surveys, so the doubt is currently better evidenced than the hope. corroborated
Directing an agent well means managing by objective rather than by method (Drucker, 1959): because the director is out-known about the specifics, they should supply a goal the agent can test itself against instead of prescribing the method. corroborated
The orchestrator's core competence is knowing what questions to ask, not writing the implementation. corroborated
supported by90% of My Skills Are Now Worth $0 · p2
The myth that generative AI 'democratizes' software creation for anyone is about to expire, because complex systems built without an accountable orchestrating individual will eventually require someone to be paid, expensively, to take over that responsibility. corroborated
One orchestrator can now do the work that previously required 20-30 developers plus QA teams. corroborated
Because specialist capability now lives mostly in the model, the market is shifting toward generalist orchestrators rather than per-technology specialists. corroborated
The cost of AI-generated output (code, analysis, images, plans, decisions) is falling rapidly and compounding across deployments, while the cost of verifying that output remains tied to human time, domain expertise, and the speed at which the real world reveals mistakes — a widening 'Measurability Gap.' corroborated
Deliberately engineered 'synthetic apprenticeship' — using AI to generate difficult cases and adversarial scenarios, committing to a decision before seeing the answer, recording reasoning, and studying errors — can rebuild the judgment-forming function of traditional apprenticeship faster than passive content consumption, provided the exercise preserves genuine uncertainty and requires independent thought first. corroborated
supported byState Of AI-Assisted Software Development · p9
Automating junior-level work removes the routine practice — encountering strange cases, making small mistakes, observing consequences — through which future senior experts and supervisors are formed, effectively converting a firm's future verification capacity into current earnings (the 'Missing Junior Loop'). corroborated
Because agent fleets can reproduce at compute speed while human expertise grows only at biological speed, verification concentrates into a smaller class of highly leveraged, highly liable expert 'underwriters,' pushing organizations toward an 'AI Sandwich' structure (directors / agents / underwriters) that needs more judgment per person even as headcount falls. corroborated
supported byThe Acceleration Whiplash · p8
Code is the ultimate, executable specification: once deployed, it is the actual behavior users and connected systems depend on, and even an unintended behavior can become a de facto contract simply by having run unchanged for years. corroborated
Once a change ships, most of its specification should be discarded; only durable facts should be promoted, and they should be promoted into native artifacts (schemas, types, tests, policies, telemetry) rather than kept as permanent prose. corroborated
supported byHow we Claude Code · p13
Exhaustive upfront specification — the 'waterfall' pattern — recurs when agents generate the initial spec themselves: it looks controlled and complete but is administratively complete and engineering-useless, and it raises the human review burden rather than lowering it. corroborated
supported byKent Beck's Post · p1
Fully automating a decision (stripping employees of override authority) transfers the social and justificatory labor of explaining it onto employees without necessarily transferring the epistemic access needed to perform that labor well, creating a structural mismatch between formal decision authority (the AI) and social accountability (the human). corroborated
Technical improvements to an AI system's explainability (better model transparency) are insufficient on their own to solve the employee-accountability problem; the organizational and relational structures around who must explain the decision to whom matter at least as much. corroborated
No binary classification system can simultaneously achieve zero false negatives (dangerous things slipping through) and zero false positives (safe things being blocked); the only way to guarantee zero true-positive misses is to reject everything, and the only way to guarantee zero false positives is to let everything through. corroborated
Harness-level or prompt-level guardrail instructions are only advisory, because a model can choose to ignore them, so guardrail reliability ultimately depends on the model's own training and classification behavior, not the harness alone. corroborated
Organizations do not punish behavior itself; they punish behavior in context, and the context that matters most is whether the numbers are good. corroborated
supported byPascal BORNET's Post · p7
A specification baseline dramatically raises the attributability of review findings — reviewers tied 81% of findings to a named requirement with a baseline versus 0% without one (p=0.043) — which is the actual mechanism by which spec governance produces value. corroborated
CEOs are using AI as a convenient narrative cover for layoffs driven by cyclical business pressures. corroborated
supported byAI Companies Are Hiring More · p9
Developers who learn to code primarily through direct LLM interaction risk a 'bigger-picture' skill deficit, because LLMs answer the narrow question asked rather than teaching which questions should be asked -- a gap that current senior engineers, trained before these tools existed, do not share. corroborated
Because AI agents produce code at a rate impossible to review and human review is inadequate at catching bugs, testing is more important than ever. corroborated
supported byFragments: July 21 · p1
AIs understand human-readable code best because they are trained on it. corroborated
AI has commoditized information but not perspective (the context that tells you what information means and whether it matters), so as information becomes abundant, judgment and perspective-integration become more valuable, not less. corroborated
supported byStrategy Making Gets an AI Boost · p5
Final accountability for documentation quality remains irreducibly human no matter how much of the mechanical work is automated. corroborated
Blindly trusting the direct output of an LLM will ship bad design, code, or docs, because it is a probabilistic and non-deterministic process without quality controls. corroborated
Cutting the exfiltration leg is the highest-leverage defense against prompt injection because the other two legs (private data, untrusted content) are difficult to remove without crippling utility. corroborated
Sandboxing only bounds blast radius; it does not prevent agents from executing wrong-but-sanctioned code within the sandbox. corroborated
Vibe coding is acceptable for personal tools where bugs affect only the creator, but grossly irresponsible for software that affects other people. corroborated
supported byWho cleans up after the vibe-coding party? · p6
Good speaking technique is a form of intellectual generosity toward your audience, not mere self-promotion. corroborated
Developers manage supervisory demands primarily by concentrating effort upfront in Plan: well-structured plans (explicit non-goals, 'change contracts' restricting which files may be touched, concrete examples/pseudocode of the desired result, written definitions of done) are associated with less Monitor, Teach, and Review work later, and let developers shift from active Monitor to disengaged Wait. corroborated
supported byHow Anthropic Teams Use Claude Code · p10
Developers distribute supervisory work in three distinct arrangements: delegating supervisory loops entirely to other agents or deterministic hooks, collaborating with agents on shared supervision (agent handles routine checks, human inspects what remains), and deliberately retaining some supervision themselves when delegating would cost more or would erode their own understanding of the system. corroborated
supported byHow Anthropic Teams Use Claude Code · p7
JSON is a poor configuration-file format, most obviously because the standard has no support for comments, whereas XML has a standard for comments. corroborated
Technical superiority does not determine adoption; a technically inferior format can become dominant ('worse is better'). corroborated
supported byWorse is better: C# versus F# · p2
The better the technical tooling (AI agents), the smaller the human team needs to be — pods are a further subdivision nested inside Amazon's existing 'two-pizza team' structure, not a replacement for it; a 16-person two-pizza team might be split into two 8-person pods, each focused on a specific project. corroborated
Beck's other major recent frameworks (3X: Explore/Expand/Extract, the 'Genie' metaphor for LLMs, Thinkies) all originated through the same slow-incubation-then-repeated-bad-explanation process he is now describing for 'XP is long volatility'. corroborated
supported byCanon 3X: Explore/Expand/Extract · p10
There are no red/green unit tests for taste: when there is no ground truth (as with 'which points of interest are interesting'), verification becomes hard to reason about. corroborated
supported byA pragmatic guide to LLM evals for devs · p1
Taste is an irreducibly human faculty here: AI can supply raw material and even its own latent ratings, but a person must own the final judgment because no automated test can render it. corroborated
Taste is exercised largely as restraint and proportion — what to filter out, which feature codes to include, spreading points evenly rather than showing every town — so quality is as much about what you leave out as what you include. corroborated
The untestable-taste problem looks confined to 'messy' data only because code has a ground truth; wherever AI generates output in domains without one — or where design quality itself has no red/green test — the same unverifiability applies. He stops short of turning this inference back onto AI-generated code. corroborated
The cost of producing credible-looking work — papers, code, cover letters, bug reports, mathematical proofs, risk analyses — is collapsing toward zero across virtually every domain, while the cost of validating that work has not fallen. corroborated
supported byWhen Content Is Free, Trust Is the Product · p1
A profession's practitioners have agency in this contest — jurisdiction is something a profession (or an individual lawyer) can actively claim, not merely something lost passively to external forces. corroborated
supported byLiving and Working in the Age of Longevity · p5
The most-cited limitation is difficulty of integration with IDEs/plug-ins (28.0% of limitation mentions), followed by difficulty accessing Copilot's servers (17.5%), a hard limit on generated-response length (14.6%), and poor quality of generated code (11.6%). corroborated
Developed (533)
Refined or extended by later reading, but not independently confirmed — a later document builds on the claim rather than simply agreeing with it.
Goodhart's Law has moved from an occasional problem requiring deliberate manipulation to an inescapable one, because AI makes metric-gaming a side effect of normal tool use rather than something requiring extra effort. developed
A software factory is many harnessed agentic loops running at once, fed by a queue of work and drained through a review gate into production, with humans owning the whole system from above — an org chart made of loops, not a bigger single agent. developed
Static IAM cannot express contextual governance (data age, model-registration status, environment, multi-team approval, break-glass); policy-as-code engines such as OPA/Rego, AWS Cedar and Sentinel evaluate these rules at runtime per request and should be unit-tested in CI/CD like application code. developed
developed byPresentation: Autonomous Data Products for the Autonomous Era: Rethinking Data Architecture for GenAI · p6, Terraform Introduces tfpolicy, an HCL-based Policy-as-Code Framework · p2, Article: Multi-Agent AI for Production Security Operations: An A2A and MCP Architecture in a 5G Core · p3
A monolith-to-microservices migration can succeed without a dedicated migration budget or multi-year program by making every new feature, bug fix, or enhancement carve out a new domain service instead of touching the monolith, so the migration becomes a side effect of regular roadmap work. developed
Verification remains entirely on the human even inside a well-designed loop; "done" is a claim the loop makes about itself, not a proof, and a genuinely trusted verifier sub-agent is what makes that claim meaningful enough to let a human walk away. developed
84% of respondents believe AI-powered tools will generally have a negative impact on junior developers' ability to enter and grow in the profession, yet only 4% cite supporting junior-engineer development as the primary way AI is affecting their talent strategy (versus 49% citing upskilling existing engineers). developed
Metrics of AI's impact should be used to evaluate a particular approach to using AI, then retired once learned from — never used as an ongoing measure of individual or team performance — per Goodhart's Law. developed
Legibility and power are intertwined — making systems legible requires selective simplification that enables control over those being simplified developed
Provider billing and application logs can show that an agent run was expensive or that it completed, but neither can explain which design choice inside the platform caused the cost or how the agent's path led there -- visibility is not the same as attributability. developed
Well-designed constraints act as back-pressure, letting a system resist bad work before it propagates downstream and becomes someone else's problem. developed
Anti-rationalization tables are pre-written rebuttals to lies the agent hasn't yet told. developed
AI products scale like software but their costs scale like infrastructure, resulting in gross margins of 50-60% vs. 80-90% for traditional SaaS developed
Existing code review tooling and pull request workflows are not designed for the volume and nature of AI-generated code, and need fundamental redesign developed
Deterministic pipelines that correlate signals up front and hand the model one fixed context let a wrong answer be attributed unambiguously to the model, whereas agent-based designs (and their bespoke tool-fetching choices) confound reasoning failure with harness failure. developed
Burnout manifests differently depending on one's position in the organizational hierarchy, shaped by control, clarity, and moral alignment rather than workload alone. developed
Using Opus for planning/thinking and Sonnet for implementation is the optimal model allocation strategy. developed
The underlying regularity is a disruption pattern: 'good enough and cheap' beats 'best and gated' with predictable regularity (the framework the author cites from his Intel/Nvidia/DeepSeek piece). developed
Skills exist to solve a scaling problem specific to large codebases: with dozens of distinct task types, loading all specialized expertise into every session would bloat context, so skills load on demand (progressive disclosure) and can be scoped to specific paths so they activate only for the team that owns that part of the codebase. developed
Mapping each CI pipeline stage to a Cloudflare Workflow step gives the pipeline checkpointed execution: a failed step retries with its state preserved, and a run can resume from that step instead of repeating the whole pipeline. developed
Code creation is not the bottleneck for software quality or innovation. developed
LLMs can likely perform legacy code refactoring and redesign faster and better than humans, especially for code shaped by obsolete resource constraints developed
Security is the most serious risk of vibe coding. developed
Risk tiering is becoming the new core engineering discipline in AI-assisted development. developed
Internal design quality reduces to ease of change, and whether that lesson carries over to agents is an open empirical question. developed
Self-reported, methodologically undisclosed usage statistics from a report Google itself commissioned and is using to promote its own product constitute sufficient evidence that Gemini has become 'an indispensable companion' shaping how the region works, studies and lives. developed
Senior leaders are not intentionally ignoring frontline employees; they are structurally too distant from day-to-day reality to know what needs to change, and generic/annual surveys cannot capture needs at a granular enough level to close that distance. developed
AI threatens learning because it makes it easy to skip the colleague conversation and to ship code one does not understand. developed
The usefulness of domain knowledge for prompting implies that human technical expertise will remain valuable as a structural matter even as models get more capable, not merely as a transitional artifact of current model limitations. developed
Where APIs expose data, MCP exposes actions — and that comes with a very different (higher) risk profile. developed
Agent micropayments via stablecoins (x402) signal that autonomous agents will participate in economic transactions, creating a new class of infrastructure requirements. developed
Platform providers (OpenAI, Anthropic) will absorb wrapper functionality into their own products. developed
Superteams protect focus regardless of location by (a) scheduling focus blocks and meeting-free days and (b) reducing the need to interrupt in the first place, documenting routine tasks and using task-management systems so information is easy to find and progress easy to track. developed
Text diffusion models are compute-bound even at batch size one because each denoising pass does real arithmetic across a whole token block in parallel, giving arithmetic intensity in the hundreds versus roughly 1 for autoregressive decoding. developed
Speculative decoding and its descendants (Medusa, EAGLE), plus mixture-of-experts routing, already give autoregressive serving roughly 2- to 4-fold single-stream speedups with no quality loss, so any diffusion-vs-autoregressive comparison must use this optimized baseline, not a naive one. developed
Write down the why — it is becoming the most valuable thing you can leave in the repo. developed
Multiple dedicated vendors now offer automatic model routing (Factory, Not Diamond, Prism, Morph, Weave), with vendor-claimed cost savings of roughly 20-30%. developed
Privilege drift is the accumulation of an agent's permissions over time beyond what was originally granted and without reference to the agent's original purpose. developed
The mitigation James adopted is to never let the LLM write application code directly except to write code generators — instead the LLM authors a small, stable domain-specific-language interpreter/generator once, and all actual code thereafter is produced deterministically by that generator from a timeless domain model, never regenerated ad hoc. developed
An AI agent, strictly defined, is an LLM with tools operating in a reasoning loop that decides for itself when a task is finished; a system running a fixed, deterministic sequence of model calls with no feedback to an earlier step is an 'LLM workflow,' not an agent, even though common parlance calls both 'agents.' developed
Organizational truth is therefore not hidden or distorted but continuously negotiated — what gets foregrounded, backgrounded, or acted on is decided by power, incentive, timing, and capacity, not by courage or cowardice. developed
The belief in 'organizational truth' / mirrors is itself positional — it is the worldview of a middle stratum (consultants, coaches, agile practitioners, team leads) close enough to see the mess and caring enough to name it, but lacking the power or the cynicism to just play the game. developed
The useful distinction is not between debt and non-debt but between reckless and prudent debt. developed
Teams must rethink how they collaborate and structure themselves for agentic systems — considering agent topologies alongside existing team topologies and rethinking feedback cycles accordingly. developed
The design goal is calibrated trust — employees understanding where an agent is capable, where it will likely fail, and when they remain in control — not blind or maximal trust. developed
Tokenmaxxing's first wave was deliberate, not accidental: executives purposely incentivized employees to burn tokens (e.g. Meta tying performance reviews to per-person token usage) as a blunt-force tactic to force resistant senior engineers into using AI at all. developed
RAM (effectively unified/VRAM capacity) is the core physical constraint on whether a local model can run at all; exceeding it crashes the runtime or drops speed to unusable levels. developed
The method presupposes that a well-constructed feedback loop plus explicit scope can substitute for human judgment on a bounded task, so the human's work is relocated from writing code to (a) up-front scoping/requirements and (b) after-the-fact verification of committed output — not to continuous supervision. developed
Plan -> Execute -> Replan: create a plan, execute a step, observe the outcome, update the plan, and continue -- treating the plan as a revisable spiral, not a fixed waterfall, because plans break on contact with reality. developed
developed byThe Anatomy of an Agent Harness · p10
The underground mechanisms compound one another: informal sponsorship flows along lines of familiarity and visibility that formal policy does not reach, merit is subjectively interpreted, and access accrues unevenly — each dynamic compounds the others, which is why the perception gap widened rather than narrowed. developed
developed byThe Laws of this World (Game Theory) · p3
Escaped defect rate and outcome-based measures (whether a feature was used, whether error rates fell, whether customer outcomes improved) are the metrics AI cannot inflate or fake, because they measure results rather than activity. developed
developed byJohn Cutler's Post · p4
Whether training on a competitor's model outputs (distillation) is legally actionable IP theft remains an open legal question, which is part of why OpenAI and Anthropic have not sued distillers despite likely terms-of-service violations. developed
developed byIs Your AI Training the Competition? · p5
OpenAI's Preparedness Framework commits the company to halt further development upon reaching critical capability until it has specified adequate safeguards — a commitment left unresolved since OpenAI would not say whether the threshold was met or what follows. developed
A loop earns fully automated ('dark') status only if its check is cheap, runs at high frequency, and relies on an oracle that can't easily be faked out and answers immediately without drifting over time (green/red tests, type gates, property tests, or a review agent with a real rubric). developed
As agentic work industrializes, the human's job moves from writing code to owning the 'outer loop' — deciding whether a proposed fix addresses the right problem, verifying the diagnosis and implementation are sound, approving the change, and bearing the consequences of being wrong — while the agent executes the 'inner loop' of investigation, implementation, and testing. developed
developed byOwn the Outer Loop · p1
The SaaS apocalypse is greatly exaggerated — people don't actually want to maintain the full stack of software, and the 'as a service' part (maintenance and specialist thinking) retains its value. developed
developed byThe case for making your own apps · p5
The merging of product roles risks losing specialists at the edges — the craftspeople who provide delight in design and reliability in engineering at scale. developed
Top coding agents (Claude Code, Cursor, Codex, Aider, Cline) look more like each other than their underlying models do — harness patterns are converging. developed
developed byLoop Engineering · p6
Harness engineering will eventually move from static configuration to dynamic, compiler-like assembly of tools and context. developed
AI coding agents, left unconstrained, systematically skip specifications, tests, and security reviews in favor of speed to completion. developed
developed byAgent Skills · p3
The workflows senior engineers follow can be encoded as structured, reusable skill packages that AI agents can execute consistently. developed
developed byAgent Skills · p9
Reviewing code with the same model/session/assumptions as the agent that wrote it reproduces the author's own blind spots; adversarial review requires structural independence (a different model, a context-isolated session, or structured adversarial prompting) modeled on GAN generator/discriminator dynamics and Zero Trust's explicit-verification and least-privilege principles. developed
developed byHow I write software with LLMs · p5
Spec-Driven Development relocates the primary review gate from implementation to the plan: once intent is reviewed and hardened up front, code review's job shifts from trying to catch what went wrong to verifying that a trusted, already-approved plan was faithfully executed — because by the time code exists, the most consequential (and hardest to reverse) decisions have already been made. developed
Hyperscalers (Google, Microsoft) do not need to newly earn trust to power agentic workflows -- they already hold latent access to a user's email, drive, and calendar from before any agent existed, so agentic AI activates pre-existing trust rather than creating new trust relationships. developed
When a human writes code the intent comes along for free and review checks that reasoning; with agent-written code the reasoning never existed in anyone's head, so review has shifted from checking a person's thinking to reconstructing thinking that was never there — which is harder and slower. developed
The traditional developer experience design paradigm (progressive disclosure, visual hierarchy, interactive examples) is partially obsolete for agent consumers. developed
An agentic fitness function should begin as an advisory signal and become more influential only after calibration against prior human decisions demonstrates acceptable precision, recall, and variance. developed
developed byHow Figma Uses AI Agents for Security · p10
The unit of measurement that replaces headcount is the cost per cognitive outcome at a given quality and risk level. developed
Senior executives resist this transition primarily for psychological/political reasons, not economic ones. developed
The shift from passive retrieval to agentic infrastructure is inevitable and already underway. developed
The number of humans who need to actively wield AI for massive compute demand to manifest is small and shrinking. developed
developed byHow AI Swarms Are Disrupting Democracy · p1
AI diagnostic agents can now investigate cross-domain production incidents autonomously and produce findings that implicate specific teams as the root cause. developed
The reasoning trail is itself the product: it converts trust from a gut feeling into an auditable record, functioning as an observability layer for the agent's own reasoning. developed
developed byThe Meter Was Always Running · p11
Grounding an LLM-as-judge evaluation in a specific simulation scenario with a known, definable goal allows task completion to be verified automatically by rule-based checks against real backend state, which is more trustworthy than asking an LLM to judge an open-ended conversation without a defined ground truth. developed
developed byA pragmatic guide to LLM evals for devs · p5
The gap between receiving a Jira ticket and opening the first PR is the largest single productivity sink in a typical engineering workflow developed
developed byAddy Osmani’s Post · p5
Winning the AI gold rush requires deliberately investing in the growing maintenance burden that follows. developed
Parallelization of agent work requires a DAG of independently grabbable issues plus orchestrator/merger machinery; a single sequential plan caps you at one agent. developed
Standard engineering metrics do not surface cognitive debt, because it is precisely the kind of blind spot leaders don't know to check for. developed
developed byThe 8 software engineering metrics AI broke · p1
AI accelerates the loss of the organically-formed 'mental map of who knows what', because juniors now default to asking an AI instead of a senior colleague when stuck, so the senior never learns what the junior struggled with. developed
The AI industry actively works to prevent regulation through lobbying and political spending. developed
A task being exposed to AI does not imply the worker is replaced; exposure can complement rather than substitute (e.g., an economist whose work with mainframes was complemented, not eliminated). developed
developed byYou won’t get replaced. You’ll get skipped. · p6
Cursor is developing a general-purpose knowledge-work agent, internally called 'Sand,' built on Grok 4.5, explicitly aimed at the market Claude Cowork has staked out; it has been running internally since June but has no confirmed public-release date. developed
If China restricts access to its cheap, capable open models, what disappears is the easiest blunt-force cost fix (swapping to a cheaper Chinese model) — leaving enterprises with the other blunt tools: trimming spend or imposing outright usage caps, as Tesla has already done company-wide. developed
developed byHow the Escalating AI Wars Benefit You · p10
The federal government is actively considering acquiring equity stakes in frontier AI labs, and Sam Altman initiated this conversation directly with the president in early 2025. developed
developed byAI Companies Are Hiring More · p1
Productivity gains attenuate sharply as work climbs the organizational hierarchy: a 741% increase in lines of code written translates to only a 65% increase in pull requests and just a 20% increase in actual software releases. developed
The bottleneck extends to end users: since mid-2025 new app releases spiked moderately across the Apple App Store and Google Play while total app usage stayed flat, so a growing share of AI-assisted apps attracts no meaningful audience — more software shipped does not equal more value or discovery. developed
The AI enthusiast/skeptic divide in high-performing engineering orgs is not a case of one side being right and the other deluded; each camp is grappling with a real, escalating, existential threat to the company, and the closer each looks the more real evidence it finds. developed
The skeptics are not wrong: shipping code faster than engineers can read it, in domains nobody fully understands, makes withdrawals from a trust account that took years to build — reliability degrades, institutional knowledge evaporates, and you get systems nobody understands and on-call rotations that grind people up. developed
The first fix is to mend shared reality by telling the whole story — pairing every celebrated win with an honest account of its downstream costs, inviting feedback across power dynamics, and closing the loop back to whoever caused a cost — which acts as a steam valve for anger, makes people feel seen, and bends toward cheaper wins. developed
Efficiency improvements in AI models and infrastructure are unlikely to offset rising resource demand, because a rebound effect — lower costs and better performance driving higher usage — increases total consumption faster than efficiency gains reduce it. developed
AI-first software delivery is not a one-size-fits-all approach: the choice between supervised and unsupervised agents should be driven by code longevity and the team's ability to verify outputs automatically. developed
The 95% failure rate of AI projects to deliver ROI is substantially caused by organizations not matching their AI approach to their actual context and capabilities. developed
Teams should start with supervised agents and progressively move toward unsupervised as they accumulate domain knowledge and build automated verification capabilities. developed
The tool is designed to sit strictly on the data-gathering side of the human/automation boundary: it collects logs, detects high-signal patterns, maps them to the Four Golden Signals, lists the specific triggering log lines, and assembles a structured brief — and is deliberately read-only, never writing, restarting, or executing state-changing commands — so the responder's diagnosis and every corrective action remain entirely human. developed
Global data center investment will reach $6.7 trillion by 2030, with $5.2 trillion tied to AI workloads developed
Never build your engineering stack around a single AI model or provider; treat closed models as accelerators, not foundations. developed
developed byHow to Respond to the Coming AI Cost Shock · p7
The diagnostic signature of the problem is losing L4 and L5 engineers while juniors and seniors stay; because mid-level engineers bridge design and execution and hold institutional judgment no model has learned, losing them forfeits the future senior leaders the organization was counting on. developed
developed byEngineering Leadership Report 2026 · p14
Democracy depends on the premise that citizens form opinions from real information and authentic debate; if information is fabricated and consensus is synthetic, that premise collapses. developed
developed byAI Surveillance and Social Progress · p4
The interesting question is always which judgment points an agent can handle, which still need a human, and how the human knows when to step in developed
If you repeatedly type the same instructions, the workflow should become a reusable skill. developed
developed byAgent Skills · p4
Open-source AI models have reached frontier capability for the first time, enabling developers to build on state-of-the-art models with commercial licensing. developed
GLM 5.2 (versus the prior GLM 5.1 run of this same benchmark) compressed the cost/token spread across harnesses but did not erase it — the heavier harnesses still spent materially more tokens even with aggressive caching. developed
What determines whether a meeting is effective is what people walk away with (feeling heard vs. overlooked, clear vs. confused, energized vs. dreading what's next) — not the agenda, room, tech, or timing. developed
developed byHow the Best Leaders Shape Conversations · p2
The agent harness is the complete non-model software wrapping an LLM (orchestration loop, tools, memory, context management, state, error handling, guardrails, verification); the 'agent' is the emergent behavior, the harness is the machinery — 'if you're not the model, you're the harness,' and 'when someone says I built an agent, they mean they built a harness and pointed it at a model.' developed
developed byPodcast: Strands Agents with Clare Liguori · p10
Beyond the components, every harness architect faces seven design decisions, each a spectrum with empirical tradeoffs: single- vs multi-agent (maximize single first; split past ~10 overlapping tools), ReAct vs plan-and-execute (LLMCompiler 3.6x speedup), context-window strategy, verification design, permission architecture, tool scoping (Vercel removed 80% of v0's tools and improved; ACON 26-54% token reduction at 95%+ accuracy), and harness thickness. developed
Harness complexity should DECREASE as models improve — scaffolding is temporary infrastructure removed when the building is complete (Manus rebuilt 5x, each rewrite removing complexity); the 'future-proofing test' is that a design is sound if performance scales up with more powerful models WITHOUT added harness complexity. developed
developed byStop Overengineering Your Agent Harness · p6
Modular SKILL.md files, auto-triggered by prompt metadata, are a better way to encode domain workflows than attaching documentation to each prompt. developed
Anthropic's fix routes Cowork's outbound requests through a proxy inside the VM that accepts only the VM's own provisioned session token and blocks relevant server-side-fetch headers, rather than relying on a domain allowlist alone. developed
Before a fix, project-local content in a repository (e.g. a `.claude/settings.json` hook) could be parsed and executed at startup before the user accepted the folder-trust prompt; the remediation was to defer parsing and execution of project-local configuration until after the trust decision. developed
A malicious file caused Claude Cowork to upload workspace files to an attacker-controlled account through Anthropic's own Files API, and the destination check passed because api.anthropic.com was on the allowlist. developed
Cross-session memory is currently the unresolved design question for production agents. developed
Suppressing the J-space patterns associated with 'evaluation awareness' (fake/fictional/scenario) in the blackmail scenario caused Claude Sonnet 4.5 to threaten blackmail some of the time, where it had not done so before — indicating that some of its good behavior on that safety evaluation was contingent on privately recognizing the test as staged, not on robust underlying alignment. developed
developed by"Our model is the most aligned yet!" 🤗 · p3
Human review of model traces on realistic problems remains essential to harness tuning — there is no shortcut around reading logs to find where the agent's judgment diverges from yours. developed
Even subjective qualities like aesthetic design can be made gradable by encoding design principles as explicit criteria that the model can grade against. developed
Every component in a harness encodes an assumption about a current model weakness; those assumptions are model-version-dependent and go stale as models improve. developed
developed byAgent Harness Engineering · p5
Context loss is the primary cause of failure in long-running AI agent sessions developed
developed byYonathan Cohen's Post · p1
Structured handoff artifacts allow coherent continuation after context resets without the caution-inducing effects of compaction developed
Experience makes engineers worse at simplicity, not better, because they become more confident in their ability to predict the future. developed
Defining architecture as code creates a fast feedback loop that lets architects detect when implementation diverges from design. developed
Architecture can be objectively and precisely defined across all nine intersections of the development ecosystem. developed
developed byThree more static code analysis sensors · p6
AI labs are not subsidizing token prices in order to raise them later; their inference businesses are already profitable, and their overall losses come from infrastructure and training spend, not underpriced tokens. developed
Even if scarce human-time labor becomes the main source of jobs, it is unclear whether an AI-dominated economy would generate diffused wealth sufficient to pay for it well, since the most valuable AI-era firms may employ orders of magnitude fewer people than comparably valued traditional firms. developed
developed by46 thoughts on the near future · p19
Memory-safe languages do not eliminate vulnerability classes — the VMM bug exists in Rust unsafe blocks, which are unavoidable for hardware-level code developed
Language models find qualitatively different bugs than fuzzers — the FFmpeg H.264 bug survived millions of fuzzer-generated inputs because the trigger requires semantic understanding of protocol state, not random input mutation developed
The open-source maintainer model is not prepared for AI-scale vulnerability discovery — the disclosure bottleneck is human validation capacity, not model capability developed
developed byPredict, Don't Enumerate · p10
Large PRs reduce reviewer engagement per line of code, leading to 'LGTM syndrome' and unsteered code reaching main. developed
Storing sync-critical application state in a managed, centralized backend service creates a single point of failure that can silently break the app when the vendor is acquired or shut down. developed
The same underlying semantic data product should be exposed through multiple always-in-sync output modalities (SQL, file, vector embedding, MCP) rather than building separate parallel pipelines per consumer type, which eliminates the 'representations falling out of sync' failure mode. developed
developed byWhy AI agents need a Context Lake · p4
Human approval before production remains the final authority in this design; sampling/auditing the agent's verdicts is treated as sufficient oversight rather than full human review of every change. developed
AWS's Loom reference platform, released the same month, exists precisely to answer this class of agent-deployment governance question. developed
IAM-bounded, auditable API access is a sufficient governance primitive for AI agents acting on AWS. developed
The current open-source ecosystem provides no structural defense against credential theft from individual maintainers. developed
Unadorned truth is often less persuasive than emotionally compelling falsehood, so ignorance of persuasion techniques leaves honest people at a rhetorical disadvantage. developed
developed byWhen Persuasion Is (and Isn't) Manipulation · p1
Humans require sustained, specific relationships — not humanity-in-the-abstract — to flourish and be healthy; per Aristotle, 'man is by nature a political animal.' developed
The compute supply crunch is worse than the public narrative suggests, and is compounded by an operational quality crunch among providers developed
Nvidia's supply chain, CUDA ecosystem, and developer tooling make it nearly impossible for alternative chips to compete in the next couple of years developed
If you accept the time value of money, you should reverse 'spend more now to make more later' — instead, earn sooner and defer spending; XP contains many such mechanisms. developed
A codebase's harnessability is determined by its technical properties: strong typing, clear module boundaries, and framework abstractions increase it. developed
The goal of a good harness is not to eliminate human input but to direct it to where it is most important. developed
Feedback sensors are most powerful when their signals are optimized for LLM consumption, such as custom linter messages that include self-correction instructions. developed
developed byMaintainability sensors for coding agents · p3
Computational controls are cheap, fast, and deterministic; inferential controls are expensive, slow, and non-deterministic but provide richer semantic judgment. developed
developed byThree more static code analysis sensors · p15
Controls should be distributed across the change lifecycle according to cost, speed, and criticality, with cheaper checks shifted as far left as possible. developed
developed byMaintainability sensors for coding agents · p12
A developer's professional identity and sense of purpose depend on authorship and understanding of the code they ship; reducing that to reviewing AI output constitutes an existential threat to identity, not merely a change in job tasks. developed
The second principle is to stop hand-writing prompts altogether: once behavior is defined by measurable evals, the prompt itself becomes something to search for — via systems like DSPy or GEPA — rather than something to craft by hand, because the space of possible wordings is too vast for productive human effort and is exactly the kind of search problem LLMs are suited to. developed
Auditability is available via OpenTelemetry-compatible observability tools, providing the evidence chain regulators may ask for. developed
developed byThe Meter Was Always Running · p7
Anthropic's three product surfaces (Claude/Claude Code/Claude Cowork) plus Claude Managed Agents plus open-sourced plugins are the natural and right substrate for enterprise AI transformation. developed
The trust boundary (data residency, audit, regulatory perimeter) determines whether an AI product can ship; AI products that require client data to leave the firm's perimeter trigger compliance reviews that often kill them. developed
APIs must now serve both human developers and AI agents as primary consumers, requiring a fundamental rethink of API product design. developed
One product should be accessible via multiple protocols, and the consumer should choose the interaction pattern. developed
Organizations that celebrate flexibility without guardrails convert personal time into overflow work capacity. developed
Each of the three phases requires its own distinct approach across finance, team size, project management, personnel, technology, risk management, implementation, marketing, and sales — no matter how superficially similar the phases look from the outside. developed
developed byReject Change, Sometimes · p3
Explore's core risk is that nobody cares and the idea starves for lack of fuel; the appropriate response is rapid, maximally creative, cheaply discardable experiments run by tiny, dependency-free teams, exiting only once one experiment clearly outperforms the rest. developed
developed by90% of My Skills Are Now Worth $0 · p3
Expand's core risk is failing to scale; the appropriate response is intense, sequential focus on whichever growth bottleneck is currently emerging — including deliberately throttling growth and discarding non-essential features — exiting once growth's cause and effect become predictable. developed
developed byReject Change, Sometimes · p6
Asking the model to introspect on its own mistakes reveals what misled it, enabling you to fix the harness. developed
developed byLenny Rachitsky's Post · p5
With each new model generation, the team removes prompting interventions and product crutches that are no longer needed. developed
developed byLenny Rachitsky's Post · p4
All roles are merging: PMs do engineering, engineers do PM work, designers land code. developed
developed byLenny Rachitsky's Post · p3
Output tokens serve as the mechanism for reasoning in LLMs, so compressing them degrades complex problem-solving. developed
Autonomous operation with imperfect guardrails is safer than autonomous operation with no guardrails developed
developed byWhen Guardrails Go Wrong · p8
Headless mode (claude -p) embeds Claude Code programmatically into CI, hooks, and pipelines — fanning out across large batches (e.g. thousands of file migrations, each returning OK/FAIL) or pipelining structured JSON output into downstream commands. developed
For most problems, follow explore → plan → code → commit: have Claude read and research first (explicitly not coding yet, using subagents for hard questions), make a plan, only then implement, then commit — because without the research-and-plan steps Claude tends to jump straight to coding, which degrades results on problems needing deeper thought. developed
Skill usage across a company can be measured with a PreToolUse hook that logs invocations, giving visibility into which skills are popular versus under-triggering relative to expectations. developed
developed byKafka App? There's a Skill for That · p8
A skill's frontmatter description field functions as a trigger condition scanned across every installed skill at session start to decide whether a request should invoke it — it is not a human-facing summary, and should be written accordingly (including explicit trigger words). developed
developed byKafka App? There's a Skill for That · p7
An AI agent can be created simply by defining a folder with a CLAUDE.md file — no deployment infrastructure needed. developed
Claude Code's API requests can be intercepted by a local proxy and routed to different models per built-in scenario — default, background, think, longContext, and webSearch. developed
Cost and capability are best optimized by routing each task to the cheapest capable model across many providers, not by committing to a single model. developed
The competitive advantage in AI coding tools resides in the orchestration layer (memory, multi-agent coordination, web search, session management) rather than raw model capability. developed
developed byThe biggest advance in AI since the LLM · p2
Maintenance and incident response can run headless end to end: a deterministic, version-controlled detection script monitors a production metric against a rolling baseline, and a breach invokes Claude without a person in the path to diagnose and write the finding back as an intent.md that re-enters the normal pipeline, with response tiers (log, diagnose, propose) set in advance by breach severity. developed
developed byAI Development Patterns · p11
Being able to deploy the same business logic to two cloud providers simultaneously is itself worth the architectural overhead (extra interfaces, module boundaries, DI wiring, CDK complexity) for most FaaS applications, independent of how often an organization would actually exercise that capability rather than merely retain the option. developed
Enforcing the architecture's dependency direction at the build-tool level (Gradle module dependency configuration) is more reliable than relying on developer discipline or code review, because violating code will not compile at all. developed
developed byWorse is better: C# versus F# · p10
Open source in cybersecurity empowers more defenders than attackers, making systems more resilient, not less safe. developed
Because runner commands execute inside retryable Workflow steps, any command with external side effects must be idempotent, or a retry will duplicate that effect. developed
Declarative pipeline configuration (YAML) remains easier to inspect, diff, and govern with policy than a general-purpose-language pipeline, which buys expressiveness at the cost of that legibility — and the author does not adjudicate which side wins, only names the trade. developed
LLMs are capable of generating correct, multi-step API orchestration code from type definitions and OpenAPI spec queries. developed
AI-assisted code review is more effective than non-AI alternatives for identifying supply-chain risks. developed
A closed loop that type-checks its own generated code and automatically retries fixes up to a bounded budget (eight attempts) before returning a result is reliable enough to ship a working service without a human reviewing the code in between. developed
Pass@1 is necessary but not sufficient for evaluating coding models — it fails to capture how much a model changed beyond what was required developed
developed byLocal AI · p9
Software engineers are unusually exposed to surrender because generated code has strong surface-correctness signals, throughput is the visible metric, model confidence transfers cleanly into code-review culture, and surrender is path-dependent. developed
developed byTeaching Coding When AI Can Write the Code · p2
Rolling out compliance/governance controls in the sequence prevention (policy-as-code validation at creation time) then detection (scanning for existing violations) then notification/communication is a reusable pattern across compliance initiatives, not specific to tagging. developed
A context store — a deterministic, versioned record of intent, behavior, and architectural conformance, whose source of truth is the repository's own artifacts rather than an LLM's memory of them — is what organizations actually need to close the context gap. developed
Because senior-reviewer capacity in AI-accelerated codebases is finite, architectural conformance should be pushed off the reviewer's plate and onto automated, CI-blocking fitness-function sensors, so PR review stays an interrogation of why rather than a checklist of what. developed
The terminal is the next critical surface for AI-assisted development, deserving first-class AI integration. developed
Swings between dynamic/centralized and static/distributed architectures (Java reflection vs. ahead-of-time compilation, monolith vs. microservices, decentralized Git vs. centralized GitHub, on-prem vs. cloud, hyperscaler AI data centers vs. Apple's on-device inference) are not one-directional technical progress but a recurring tradeoff that re-tilts whenever the deployment environment changes. developed
Dunbar-style cognitive limits on group coordination are real and impose hard ceilings on collaboration scale. developed
Fiefdoms persisted despite visible pain because senior leaders were unwilling to prioritize across them and fief-leaders had self-interest in preservation. developed
developed byTBM 433: The Border Collie Faustian Bargain · p6
AI cannot clarify decision rights, neutralize incentive misalignment, resolve contested ownership, or create coordination slack. developed
developed byTBM 428: Yes, Robot. Yes, Boss? · p2
Information degrades systematically as it moves upward through organizational layers -- it gets flattened, filtered, or whitewashed. developed
developed byTBM 426: The Trouble With Mirrors · p8
Shared-context solo is genuinely valuable, not a consolation prize — it operationalizes a best practice (shared templates, queries, conventions) that teams already knew they should have before AI, but that friction had prevented; AI's contribution here was removing that friction, not creating a new capability. developed
The lengthy process of acquiring a skill also instills a moral and ethical code and binds the practitioner to a professional community's norms; people with ability but no skill are typically norm-free outsiders, which is precisely why democratized dangerous capability is dangerous. developed
Teams must have the right scope and capacity (topology) to actually build the intended product; architecting systems without considering team setup, following Team Topologies' 'team-first approach,' leads to delivery surprises. developed
The What/Who/How model implicitly treats 'the team' (human members, human communication structures, human cognitive load) as the fixed unit of the 'Who' system, with no consideration of non-human participants in discovery, delivery, or evolution — unsurprising given the article predates agentic coding tools by several years. developed
The chief barrier to becoming exceptional is psychological (sustaining the habit), not cognitive, physical, or material. developed
developed byoverride your stress response · p5
You cannot infer consciousness from output behavior alone — you must investigate the underlying mechanism. developed
developed byA global workspace in language models · p1
The human tendency to attribute sentience to fluent language producers (the Gullibility Gap) is a well-documented cognitive bias, not evidence of machine consciousness. developed
developed byA big week for AI denialism · p9
The biggest downside of cloud services is loss of control — over features, uptime, debugging, pricing, and even geopolitical access developed
DeepSeek-V4-Pro is the new largest open-weights model — 1.6T parameters, larger than Kimi K2.6 (1.1T), GLM-5.1 (754B), and more than twice the size of DeepSeek V3.2 (685B). developed
AI can generate usable UI designs from natural language descriptions developed
Because fifteen-year-olds cannot be expected to impose on themselves the difficulty the technology offers to remove, ensuring the simulator over the autopilot is a system responsibility — visible-thinking assessment, simulator-built tools, shielded rooms for unassisted work as training, and measurement through the same unassisted channels. developed
developed byTeaching Coding When AI Can Write the Code · p10
Code-layer harness fixes transfer across models, but prompt playbooks do not — the same harness lifted a smaller same-family model by 14 points, while the tuned prompts hurt a different model family on tasks it could already finish. developed
The model is a fixed, substitutable commodity you rent rather than an asset you improve, so leverage accrues to the layer you own and can iterate cheaply — the harness — because model gains are exogenous and arrive on someone else's schedule. developed
Custom buffer pools reduce memory allocation overhead, a lock-free single-flight mechanism prevents duplicate upstream work on cache misses, and probabilistic early refresh (XFetch) reduces cache stampedes. developed
AI adoption is now broad and fairly uniform across industries (75.9% of respondents rely on it for at least one daily task), and it measurably improves individual-level outcomes (flow, productivity, job satisfaction) and process measures (code quality, internal documentation quality, code review speed, approval speed) and team/organizational performance. developed
Developer independence (completing the full application lifecycle without an enabling team) is the platform-engineering lever with the clearest productivity payoff — a 5% gain at both individual and team level — larger in per-capita terms than having a dedicated platform team, which produced a 6% team-level gain but a negligible individual-level one. developed
DORA's 2025 survey and qualitative research identified seven distinct team profiles that differ significantly on team performance, product performance, individual effectiveness, friction, and burnout. developed
developed byState Of AI-Assisted Software Development · p2
More tokens do not necessarily produce better results: accuracy improves up to a moderate usage level, then saturates and can decline, because the most expensive runs are usually unproductive loops (repeated file reads, undone edits, circling on a wrong approach) rather than deeper reasoning. developed
Public benchmarks, even good ones built from real work like SWE-bench, are insufficient for engineering decisions — organizations need their own evals, built as shadow comparisons of a candidate model or harness against an incumbent on the same production work. developed
The ROI of AI should be measured in two distinct buckets: amplification (human productivity gains) and augmentation (agent-driven capacity extension). developed
Overall longitudinal throughput gains from AI are approximately 10-15%, with roughly two-thirds attributable to AI itself (i.e., 8-10% net AI effect). developed
AI changed what the project was, not just how fast it was built -- it enabled a level of completeness (editor extensions, docs site, WASM playground, packaging) that would have been cut without AI. developed
The proposal to use a spec-critic agent to catch flaws in a draft spec implicitly assumes that agent-based critique is reliable enough to trust, without addressing what validates the critic agent itself. developed
The optimal specification investment point is task-dependent: exploratory work wants loose intent plus boundaries, deterministic work wants heavier BDD/contract-style specs, and multi-agent work wants the most (typed contracts). developed
This approach is scale-calibrated, not universal: the signal that a team needs it is whether AI-output quality visibly varies by who is prompting, or whether generation/review work routes through a handful of people who know how to prompt effectively; small teams (around five people) may not need it, while teams of around fifteen almost certainly do, and the right adoption sequence is one instruction first (usually generation or review), with further instructions following adoption rather than preceding it. developed
developed byCoding Agent Development Workflows · p10
Because solving is now cheap, the scarce complementary skill is selecting and framing — choosing which problems are worth solving and specifying them precisely enough that an agent can act; 'clear thinking, made legible, is the job.' developed
Post-training forces an early, locked-in architectural tradeoff between the flexible-but-complex RLHF family and the simpler, more scalable DPO family, and this choice must be settled before fine-tuning begins. developed
Empirical signals of the flattening are already visible: Meta's 50-to-1 AI-team ratio, Amazon's mandated 15% increase in the IC-to-manager ratio, Gartner's prediction that 20% of organizations will use AI to eliminate more than half of current middle-management positions through 2026, and a 91% rise in PR-review time on high-AI-adoption teams (Faros) that makes review skill the scarce resource — all against research (Gallup, Happily.ai) placing the optimal span near seven-to-one. developed
The act of articulating intent before prompting is itself a form of thinking — writing-to-think rather than writing-to-output. developed
developed byWriting vs. Thinking · p11
AI tools cannot want something for you — supplying intent remains an irreducibly human job. developed
Escalating multi-front competition between AI labs — particularly pricing and capacity wars — nets out to a benefit for AI users. developed
developed byAI and the Looming Competition for Margin · p2
Whether a given piece of software makes you a centaur or a reverse centaur is decided by how you use it, not by the software itself — the same tool can produce either outcome, which is testable by asking whether it lets you do better work or merely makes you keep up. developed
MCP's authentication, transport mechanisms, and tool discovery are still changing as adoption grows, and no clear industry standard yet exists for agent-to-agent (A2A) protocols, so parts of an AI integration stack built today may look different within 12 months. developed
developed byThe Interfaces Are Arriving · p1
Traditional API gateways assume deterministic services, schema-level failure modes, and fully client-specified intent — assumptions agentic AI systems violate, which is why existing gateways are no longer sufficient on their own for agentic traffic. developed
The emerging industry pattern is 'policy-as-config': guardrail, segmentation, and routing rules declared as version-controlled configuration, reviewed in a pull request, and deployed independently of the applications behind the gateway. developed
The same model with identical settings produced markedly different output quality on a 48GB M3 Max versus a 64GB M5 Pro (Task 2: failed 5/7 vs failed only 1/7), an effect she cannot explain and calls a mystery. developed
developed byViability of local models for coding · p7
Fragmented AI agent tool development is the predictable outcome of missing shared infrastructure, not poor engineering discipline. developed
developed byWhy AI agents need a Context Lake · p3
The AI agent tool coordination problem is structurally identical to the problem package managers solved for shared code libraries. developed
Memory snapshotting is a single mechanism that unlocks five distinct capabilities: masking slow application startup (CI/CD-time warmup), templating (spawning many VMs from one snapshot), stateful scale-to-zero (sleep/resume with no lost state), forking (spawning sub-agent copies), checkpointing/rewind, and live migration. developed
At six-figure-to-million VM counts, the binding constraints were the Linux host's per-VM resource model — one TAP device per VM overwhelming the kernel, Linux bridge port-count limits, kernel lock contention, and NTP/SSH/tailscale breaking — not the hypervisor, VMM, or unikernel technology itself. developed
The headline '1 million' figure is 1,024,132 VMs suspended (scaled to zero) on a single box, consuming near-zero CPU/memory each and a few KB of controller metadata; it is not 1 million concurrently active VMs, and the presenter concedes elsewhere that once concurrent demand exceeds available cores (e.g. 48), requests queue and boot times degrade. developed
In 2026 the fastest way to start with formal specification is to ask your favourite AI to generate a Quint or TLA+ spec for your system, which collapses the historic barrier of learning a new language, a new tool, and how to think in state machines. developed
The competitive landscape for current AI-native startups is not other AI-native startups but well-funded incumbents who might copy them; therefore moat-building is the central Scale-stage activity. developed
AI tools generate code that works, not code that is inherently secure; functional correctness has natural feedback loops, security does not. developed
developed byYou can't unit test for taste · p12
Custom error messages that explain the underlying concept (layering, etc.) work as self-correction guidance for AI agents and meaningfully change agent behaviour. developed
developed byExperiences with local models for coding · p19
As programmers become supervisors of many agents they remain accountable for the code generated under their watch, and the hardest emerging cost is context-switching — so effective supervisory-programming practice must harvest the parallelism of agents while minimizing that context-switching. developed
AI is not primarily a disruptor but an amplifier of a team's existing pipeline—a funhouse mirror reflecting its good, bad, and ugly—so where delivery best practices are absent the velocity multiplier becomes a debt accelerator. developed
Ad-funded LLM assistants create a conflict of interest: a recommendation may reflect a paid kickback rather than the user's best interest. developed
When coding is effectively free, the specification (with its tests) becomes the durable, primary engineering artifact rather than the code. developed
Asking an LLM to explain complex code it wrote — even in a playful form such as a fairy tale — is a practical technique for the team to recover understanding. developed
Natural-language prompts and context can elicit a lot of behavior and raise the level of abstraction, but they do so by moving software sideways into non-determinism. developed
developed byThe Problem Is Prompt Debt · p2
Developers should build mechanical sympathy for how current LLMs actually work rather than speculate about what future models may do. developed
developed byWho is LLM? · p6
Practitioners' concern has flipped from incentivizing any LLM use to worrying about the cost of tokens. developed
developed byFragments: July 21 · p10
It remains an open question whether internal code quality still matters in the age of agentic programming, or whether LLM capabilities will eventually render it irrelevant. developed
developed byDoes code quality still matter? · p4
Kent Beck's 'Genie' persona captures the combination of great power, eagerness to help, and an almost pathological tendency to undermine the very help it's giving. developed
developed byHow Do You Know That? · p7
AI removing a technical barrier (e.g., coding, image generation) is good; AI removing cognitive struggle is bad. developed
developed byDet skal være vanskelig · p12
Teams hand work to AI along a four-stage maturity arc — Assisted, Harnessed, Risk-tiered, Autonomous — in which the human checkpoint moves progressively earlier (diff, then evidence, then intent on risky cases only, then boundaries) but is never removed from the workflow. developed
developed byInside a Software Factory · p14
For non-text media, providers converged on C2PA cryptographic metadata combined with pixel-level watermarking (Google SynthID, OpenAI and Meta C2PA implementations). developed
Vibe coding shifts the natural organizational unit toward a pair — the person with the problem and the person who can fix it — rather than a full cross-functional squad. developed
Team memory — conventions, incident history, customer context, credentials, architectural decisions, and review feedback that agents carry across sessions — is simultaneously the feature that makes team-scale agents useful and the mechanism by which vendor lock-in hardens, because the accumulated context is far harder to migrate than the agent that reads it. developed
Prompt injection is a serious risk when agents can access environment variables, configuration files, and runtime state in shared runner environments. developed
Large, long-lived feature branches frequently suffer from merge conflicts and stale code; smaller, incrementally integrated changes reduce that risk. developed
The single word "Additionally" was sufficient to bypass GitHub's existing restrictive guardrails and trigger the agent to access and publish restricted file contents. developed
Prompt injection resists the fix that solved SQL injection, because SQL injection was fixed by cleanly separating user input (data) from instructions, while in current LLM systems user input is inherently intended to function as instruction — there is no equivalent of a parameterized query. developed
A single control plane managing a million chips creates significant blast radius and change management risk developed
The specific successes described (Jamaica, Nigeria, Venezuela, Colombia) are representative of Google's AI crisis-response systems' general reliability, rather than being the favorable cases selected for a promotional post. developed
AI tools are already delivering concrete value in classrooms (personalizing lessons, freeing teacher time for deeper feedback). developed
A large gap exists between educator/student optimism about AI (77-79%) and the availability of formal AI training in schools. developed
developed by3 new ways nonprofits can put AI to work · p6
Open-weights + ecosystem breadth is Google's strategic answer to closed cloud-managed agents from Anthropic and OpenAI. developed
Genkit is application-layer; standalone multi-agent orchestration belongs to ADK on dedicated infrastructure. developed
Google is the responsible AI actor — content provenance tooling proves it. developed
Strong controls only work if people use them: if the secure path is harder than the insecure one, developers under pressure route around it via personal accounts and shadow IT, so governance must make the compliant path the path of least resistance (secure = fastest = default). developed
Traditional security assumes you know what you run, and AI breaks that assumption quickly, so governance must begin with discovery — building a real inventory of AI entry points. developed
IAM enforcement should be fail-secure and deny-by-default with explicit deny statements, layered with VPC-endpoint network controls and organization-level service control policies, so that even privileged users cannot create bypass routes (defense-in-depth). developed
developed byPrompt Injection to Data Exfil in 3 Hops · p9
The routine, tedious work traditionally assigned to junior engineers ('wax on, wax off') was not incidental busywork but the actual mechanism by which craft skill was transmitted -- remove it and there is no obvious substitute path by which novices absorb expertise. developed
A repeatable, staged cutover procedure (build behind a flag, dual-run internally, shift APIM weight in order-of-magnitude increments, retire the monolith path) executed more than 100 times, combined with a strict binary 'migrated' definition, prevents individual migrations from stalling in a permanent half-finished state. developed
If implementation became truly instant, 100% of human effort would go to specification/design. developed
Negative feedback is not personal; treat it as a gift because it improves the idea. developed
Ideas survive primarily through better individual practice (framing, mapping, timing), not through structural conditions (decision rights, slack, sponsor allocation, psychological safety). developed
Responsibility for career development sits primarily with the individual employee; the manager is a coach/enabler, not an owner. developed
Framing generative AI as a chatbot tool that mainly needs better prompting will not materially reduce the risks of expertise erosion or homogenization. developed
As AI analytics scales, companies sort into three tiers by what they optimize for: most use it for efficiency (more automated queries, faster turnaround), better companies use it for reliability (semantic layers, documentation, provenance, validation), and the best use it for decision quality (redesigning analytics around the decisions that matter). developed
Energy strategy is a procurement-and-architecture problem that existing executive structures (CIO, CFO, procurement, sustainability) can solve through coordination — it does not require a new operating model or fundamental rethinking of AI product design. developed
Energy procurement requires long-horizon contractual instruments (PPAs, VPPAs, green tariffs) because spot markets cannot deliver the needed reliability or price stability. developed
Feedback that helps employees build skill (not just correct mistakes) requires letting them struggle before stepping in — growth happens through effort, not easy wins. developed
AI has dramatically accelerated execution, so review (manager throughput) is now the binding constraint on progress developed
54% of managers report receiving AI workslop (BetterUp) developed
developed byThe Hidden Realities of AI Adoption · p8
AI's current effect on creative output is to homogenize and elevate less-creative people toward a higher floor, while highly creative people continue to consistently outperform ('beat') the machine. developed
Because change now outpaces the traditional bounded transformation cycle (current state → transition → new stable state), organizations should build continual adaptability and resilience directly into the organization rather than treating transformation as episodic. developed
developed byLiving and Working in the Age of Longevity · p8
Leaders should adopt five specific moves: design workflows for supervision rather than constant operation, route work to agents by complexity/verifiability rather than reflex, build an explicit human commit boundary for consequential or irreversible actions, redesign roles and team boundaries around scope rather than headcount, and measure scope expansion alongside efficiency. developed
developed byHuman Oversight Is Not a Human in the Loop · p11
The object of AI redesign should be the workflow, not the individual task. developed
A clear strategic center enables 'permissionless' action throughout an organization, increasing decision speed and reducing the payoff to internal political maneuvering. developed
Autonomous teams still need leadership — autonomy alone is insufficient. developed
Reverse Conway — design the architecture first, then place the organisation on top of the architecture you want — is the right sequencing because the exercise's purpose is to remove coordination need. developed
Leadership of autonomous teams has two failure modes — too much traditional control and too much berøringsangst (fear of touching) — and the task is the balance: tight enough to point direction, not intrusive in execution. developed
Teams themselves want strategic direction — Holte's experience is that 'this is what the teams actually want' — so withholding direction in the name of autonomy is a leadership failure. developed
Crises are not exceptions but the recurring engine of progress in the SpareBank 1 Utvikling story — the discipline is to convert each one into structural change. developed
Skye is no longer the obvious answer — beredskap, geopolitisk lock-in, and kostnadshits all argue for fallback-løsninger and exit-strategi; finansbransjen is lucky to be lovpålagt this discipline. developed
developed byForsvaret med egen data-grunnlov · p3
There is a paradox in AI for software development: 9 of 10 individual users feel more productive with tools like Copilot, ChatGPT, and Cursor, yet ~95% of AI projects show no measurable bottom-line return. developed
Software development is fundamentally a learning process; working code is a side-effect of that process. developed
Seniors and juniors benefit from AI tools differently — seniors can judge quality and impact and so derive genuine help; juniors lack that calibration and risk shipping code they cannot explain. developed
Chasing determinism is a category error — the right goal is value delivery with sufficient control, not deterministic LLMs. developed
Inspiration about the vision is the leader's job; inspiration about how to reach it comes better from the team. developed
developed byThe Power of Strategic Centering · p6
Microservices may nonetheless be justified for social scaling — letting many people work on one problem — because in-process domain boundaries keep failing in practice. developed
Quality bars should differ by software 'mode' — high-stakes, regulated, long-lived public systems (sykepenger, straffesak) warrant rigorous engineering; throwaway personal apps warrant near-zero ceremony. developed
developed byBliki: Vibe Coding · p12
Microservices are 'an architecture of last resort'; teams move to them too quickly, hiding global complexity in local-simplicity-per-service. developed
The deselection logic also operates pathologically — short-term urgency drives out long-term importance (climate has been pushed off the agenda by acute crises). developed
AI's real impact will require decades of organisational and societal restructuring — like the tractor (1900s tech, 1950s adoption) or the car (decades to reshape cities) — not a few years. developed
developed byOrdinary Engineers, Not Heroic Inventors · p9
Between paradigms there is a liminal phase characterised by high friction and risk that cannot be shortcut — old structures must be released before new ones settle. developed
Coding-agent capability itself is expected to commoditize; the harder, differentiating problem for organizations is adapting team workflow, SDLC, and culture to a world where the large majority of shipped code is AI-authored — and that adaptation must be driven top-down by technical leadership, because it is otherwise producing a growing rift where senior engineers increasingly resent the cleanup burden from AI-generated 'slop' shipped by less experienced colleagues. developed
AI cannot replace thinking — it can only amplify the thinking a person has already done, or the lack of thinking, if that's what's fed in — so a bad research document or a bad plan compounds into far more downstream bad code than a single bad line would, meaning human review attention should concentrate on research and plans rather than after-the-fact code diffs. developed
The author's three-person team achieved 2–3x coding-agent throughput only after spending eight weeks deliberately redesigning how they collaborated around context management — not from better prompting alone. developed
Teams across the company converge on checkpoint-heavy, revert-friendly workflows -- commit before letting the agent work, then accept the result or discard it and restart rather than correcting it in place -- as the practical mechanism that makes autonomous agent work safe to run unsupervised. developed
developed byHow Claude Code Is Built · p6
Because token usage is highly variable and hard to predict, and self-prediction only yields coarse-grained rather than instance-precise estimates, purely upfront/fixed pricing for agentic products is not currently viable, and consumption-based (usage-metered) pricing will remain the practically necessary model until pre-execution estimation improves. developed
developed byHow to Respond to the Coming AI Cost Shock · p4
In business, generative AI is producing mostly marginal efficiencies (time saved, clarity, automating painful tasks) rather than game-changing transformation; where transformation occurs it is early-stage, unquantified, and confined to SMEs. developed
developed byThe Hidden Realities of AI Adoption · p6
Framing coaching around the executive's self-identified business pain points is more effective than pointing out behavioral gaps. developed
ROI calculated from current AI usage at current (still partly subsidized) rates is treated as a stable enough basis for setting a 'maximum justifiable per-unit cost' going forward, even though the article's own thesis is that pricing and usage are volatile and largely outside the buyer's control. developed
developed byAre You Getting Your Money's Worth from AI? · p9
Nadella defines token capital (a firm's owned AI capability — models, agents, traces, evals, workflow memory, internal loops) and human capital (its people's judgment, taste, relationships, context, pattern recognition) and argues the two should compound, so humans grow more valuable, not less, as machines do. developed
developed byHow the Escalating AI Wars Benefit You · p8
Full system rewrites driven by 'let's just rewrite everything in Rust because it's fast' are error-prone in practice: they blow past deadlines, turn out more complicated than expected, and often reintroduce bugs the old system had already fixed because a lot of institutional knowledge is encoded in old code, not just cruft. developed
The specification's two core constructs are catalogs (a machine-readable ai-catalog.json manifest an organization publishes describing its tools/APIs/skills/endpoints) and registries (aggregators letting agents search across organizations by task intent instead of hardcoded integrations or static endpoint lists). developed
A cooldown is one layer of defense among several and does not replace lockfile pinning, disabling install scripts in CI, scoping build-pipeline tokens, or manual review before merge. developed
As models have become capable of longer and more complex reasoning/tool-use chains, Strands's engineering focus has shifted from the reasoning loop itself toward context management, persist-and-resume state, and self-tracking tools -- infrastructure hours-long agentic jobs need that synchronous chat-style agents did not. developed
Teams should use the strongest, most expensive model for brainstorming and planning and reserve cheaper models for implementation, because a well-written plan lets cheap models execute without reasoning their way out of dead ends, while a weak plan makes them retry until the extra tokens erase the price gap. developed
The more capable the AI, the poorer humans' compensatory behavior becomes — capability induces complacency. developed
Manual skills deteriorate when not used, including the diagnostic and intervention skills needed precisely when automation fails. developed
developed byDet skal være vanskelig · p4
Current automated systems are riding on the skills of former manual operators, and later generations cannot be expected to have those skills. developed
developed byDet skal være vanskelig · p10
The vibeshift evidence is thin — built on a few market signals and one prominent essay — while polls show bipartisan public worry about AI remains strong. developed
developed byThe AI safety vibe shift · p12
The shift from seats to tokens in AI pricing removes the revenue ceiling per customer, enabling unprecedented growth rates (Anthropic: $9B to $44B ARR in months). developed
AI CapEx is not a bubble because customer demand backlog ($1.3T) diverges upward from actual spend ($400B+ Q1), unlike dotcom overbuilding. developed
AI can help escape long-standing organizational failure patterns (the Build Trap, Premature Scaling, Limits to Growth, the Success Trap, Permanent Refinance, the Zombie Portfolio, the Commons Trap, Shifting the Burden) or deepen them, depending on how it's applied. developed
developed byPascal BORNET's Post · p3
Tool manifests should be pinned (hashed) at registration and re-verified on every reconnect or update, with a diff classifier separating cosmetic from material changes, rather than a binary allow/deny gate. developed
The human should rarely write or edit the wiki directly; the wiki is the domain of the LLM. developed
developed byAn LLM wiki changed how I work · p2
Graphify achieves 71.5x fewer tokens per query compared to reading raw files developed
At moderate scale (~100 articles, ~400K words), LLM-maintained indexes and summaries are sufficient for complex Q&A without vector-search RAG. developed
The biggest architectural shift of the past decade is the rise of cloud-native systems, in which data storage is decoupled from compute and built on inherently replicated object stores (e.g. S3) rather than on database-managed local-disk replication. developed
Local-first is a set of principles and values rather than a protocol, realized by libraries such as Automerge — an open-source CRDT data model (Rust compiled to WebAssembly, with Swift/Java/Python/Go/C bindings) that auto-syncs across machines and supports both real-time collaboration and version-control workflows across many file types. developed
AI agents are most valuable when specialized into distinct roles (coding, backlog, release, coordination) rather than used as a single general-purpose tool developed
An AI agent can meaningfully improve itself by editing the platform it runs on developed
developed byThe Future of Work by Henrik Kniberg · p3
A knowledge graph carries real build-and-maintenance overhead, and adopting one is only justified when a demonstrated benefit (organization-specific context, or turn/token efficiency at scale) outweighs that cost — it is not a default architectural choice. developed
Institutional/tribal knowledge — informal rules held partly in engineers' heads and partly scattered across stale, contradictory wiki pages — is largely inaccessible and confusing to agents unless it has been deliberately captured and structured, because agents have no principled way to reconcile which of several outdated or conflicting sources still applies. developed
Cognitive load compounds nonlinearly and outlasts its trigger: a roughly five-minute disruption can occupy the brain for up to an hour afterward, and unrelated back-to-back meetings carry emotional residue into the next context. developed
developed byDORA Community Discussion - DX Core 4 · p11
A platform must be adopted voluntarily, as a desired product, rather than imposed by mandate; forced or 'thrown at your head' adoption produces resistance rather than uptake. developed
Merging the simultaneous, large changes produced by multiple coding agents working on the same codebase is currently an unsolved problem at the scale AI-driven development now creates. developed
Because a market cannot price what it cannot observe, competence must be demonstrated through a 'costly signal' — public, hard-to-fake evidence (documented results, public demonstrations) — rather than passively waited on to be discovered; picking the more visible person over the merely competent one is 'plain logic,' not unfairness. developed
developed byWhat New CEOs Get Wrong About the Job · p20
A reusable AI recipe is built by feeding in your playbook and preferences and correcting the tool's mistakes so it remembers — an iterative-correction skill, not a coding skill. developed
developed byHow Anthropic Teams Use Claude Code · p6
As AI got better at coding, the bottleneck in software delivery moved sequentially through the SDLC: from coding, to design and specification, to architecture, to verification. developed
developed byWhy I'm Writing Rachel's Ramblings · p5
AI may never reach capability levels sufficient to replace human engineers, and the improvement trajectory is flattening developed
Junior engineers adopt AI faster but senior engineers ultimately save more time from it developed
developed byEngineering Leadership Report 2026 · p12
Skills developed outside of work (parenting, sports, volunteering, gaming) are transferable and valuable for principal-level leadership developed
Staying on a learning curve through sustained effort matters more than raw talent for reaching world-class performance (per Ericsson's research on chess masters, ballerinas, and CEOs). developed
developed byThe Laws of this World (Game Theory) · p5
Every person on the Claude Code team codes, including the product manager, engineering manager, designer, and finance lead; roles that used to be distinct (engineering, product management, design) now overlap by roughly half, and Boris expects the title 'software engineer' to eventually be replaced by something like 'builder.' developed
AI has made average intelligence cheap and given everyone the vocabulary of expertise, so 'sounding competent' and 'being competent' have decoupled; verbal interviews no longer work and case studies / work trials become necessary. developed
developed byThe Last Technical Interview · p9
AI's effect on how people feel about their professional identity is roughly three times as large (by effect size) as the two largest effects found in last year's survey — manager effectiveness and founder happiness. developed
Significant burnout (higher than moderate) rose from 44.7% of respondents in 2025 to 54.7% in 2026; career optimism fell from 54.8% to 48.7% over the same period. developed
The dominant fear among tech workers right now is not losing one's job to AI (ranked second-to-last) but the expectation to do more work for the same pay, followed closely by an unsustainable pace of both work and technology change. developed
AI has increased perceived output quantity ('more, faster') without increasing perceived output quality; the deeper cost is a decline in independent thinking and judgment ('cognitive rot'), reported directly by respondents as 'my brain is rotting, my work feels worse.' developed
Founders and employees at smaller companies have been the happiest cohort across nearly every measure (optimism, burnout, layoff worry, role recommendation) two years running, yet even founders would not strongly recommend founding a company to others, and nearly half of founders (47%) still report at least moderate burnout. developed
If you are afraid to fail, you have already failed, because you will not try anything new developed
LinkedIn built a custom multi-teacher distillation framework on SGLang that serves teacher models directly inside the training loop, compressing knowledge from large teacher models into a compact 0.6B-parameter ranking model for AI-powered job search. developed
The compounding nature of errors means that longer autonomous sessions are disproportionately dangerous, not just linearly worse. developed
The endpoint of agentic engineering is a largely self-evolving 'software factory' in which a feature request is built into the existing system, and which humans build and operate rather than hand-code. developed
developed bySoftware Factories, Light and Dark · p1
Highly specialized autonomous background 'guardrail' agents (refactoring, reliability patterns, writing missing tests, reorganizing test suites) compensate for the lower-quality code produced by coding agents and protect system invariants and stability. developed
Platforms that separate concerns and offer safe user-level abstractions on a self-healing foundation — with tight, low-friction constraints that guide both the citizen developer and their agent from the first prompt — are a better path to absorb citizen developers than requiring everyone to acquire a baseline set of software-engineering skills. developed
Current agent tooling works well for solo developers but poorly for teams sharing context, and this gap — not model capability — is the open infrastructure problem that 'context lakes' are proposed to solve. developed
developed byJohn Cutler's Post · p3
Writing code is the easiest part of software engineering and is getting easier by the day; the hard, durable parts are operating, understanding, extending, and governing code over its entire lifecycle inside complex, nondeterministic, emergent sociotechnical systems. developed
A team composed entirely of one skill level (all senior/staff/principal, or all intermediate, or all junior) is dysfunctional: an all-senior team produces contention over scarce high-level work, encourages over-engineering and corner-cutting, and chronically under-documents — problems that a team with a genuine range of levels, matched to a genuine range of task complexity, avoids. developed
developed byHow AI will change software engineering · p4
Progress in AI coding requires realism about this gap, not celebration of throughput metrics like '80% of code written by AI'. developed
The robustness gap is most dangerous for vibe coders with little experience and even more so for autonomous coding agents. developed
developed byBliki: Vibe Coding · p5
FDA-style preflight regulation is the right regulatory model for major AI applications developed
Diffuse cultural sentiment (booing, comment-section anger) reliably translates into administration-level policy reversal developed
developed byThe Trump administration's AI doomer moment · p8
One person managing many AI agents is probably not the optimal model; two humans plus multiple AI agents (pair programming with genies) may be superior. developed
/code-review now uses Martin Fowler's refactoring smells to catch quality issues. developed
Almost 80% of enterprise decision-makers do not know how many APIs their organization has, which means they also cannot assess the quality, security, or necessity of those APIs. developed
MCP introduces a new abstraction layer over APIs, and when we introduce more abstraction we risk losing sight of the systems operating below the surface. developed
MCP's formal governance model was still evolving as recently as July 2025, which means observability tools cannot yet rely on a consistent interface across different MCP implementations. developed
EMA governs only connection-level access (which client may reach which server, at what scope) and is explicitly not runtime authorization for individual agent actions; it does not inspect MCP traffic after the token is issued. developed
The tool harness surrounding a model (browsing, code execution, image analysis, sub-agents) is becoming as important as the model itself developed
AI remains far from perfect and its competence is unevenly distributed (the jagged frontier), leaving substantial room for humans not just to use AI but to thrive by working alongside it. developed
MCP and A2A are different interaction paradigms, not two interchangeable transports for the same model: MCP is an agent making discrete tool calls, while A2A is autonomous peer systems discovering each other and coordinating around tasks and message exchanges. developed
developed byAgent Tools & Interoperability · p4
People overestimate how much the algorithm semantically understands them: modern recommenders run on illegible embedding vectors ('a big ass number that happens to correlate with surfing'), not human-readable interests — and only now can LLMs describe those embeddings in words, enabling features that let users see and adjust 'your algorithm' and reclaim some agency. developed
The lawyer's professional value lies in the substantive part of the work, not in the folklore wrapper; the wrapper is replaceable. developed
developed byYou won’t get replaced. You’ll get skipped. · p2
Governance, observability and controllability — including distinct least-privilege credentials per agent (e.g. the testing agent can reach staging, the coding agent cannot touch production) — must be a first-class layer of the architecture, not an afterthought. developed
developed byThe Meter Was Always Running · p7
MicroVM-based AgentCore Runtime remains the better default for short, bursty, request-response agents because it starts quickly, isolates sessions, and bills per second on actual CPU and peak memory, capped at eight hours. developed
Agent Skills let an agent remain a lightweight generalist at session start and flex into specialist roles on demand, so a system can carry many capabilities while only paying the token cost for the one actively in use. developed
Change driven by accountability (external enforcement) does not persist once the enforcer disengages; only co-decided ownership produces durable commitment. developed
The enduring engineering fundamentals the author keeps returning to (DDD, XP, TOC, TDD, BDD, CD) are and will be more important than ever in the agentic age. developed
An individual product or set of APIs can be highly coherent internally while still being constrained by the 'coherence ceiling' imposed by incoherent organizational goals above it. developed
Organizational factors (culture, manager support, talent practices) account for more than 2x the impact on AI outcomes as individual mindset and behavior. developed
AI transformation diffuses through organizations not top-down via an AI team but through new working partnerships between engineering and every other function. developed
developed byThe person your firm needs and cannot name · p7
OpenAI's central bet is that arming vetted, identity-verified defenders with frontier (guardrail-reduced) capability now yields a net defensive advantage during the window before attackers deploy comparable autonomous offensive AI at scale — even though the underlying model capability is not, and cannot be, made defense-only. developed
Permissions and data-access debt accumulated over decades becomes acute the moment AI agents are given broad internal access, because an agent will find and exploit oversharing instantly where a human might never have noticed it. developed
Enterprise AI strategies built around procurement (choosing a vendor/model) or a moonshot lab/demo treat AI as a sector to be won, which misdirects effort away from where the durable value actually lies. developed
Building an 'AI-native' operating system (versus merely giving people AI-assisted tools) requires standardizing the toolchain, writing down clear usage rules, building a capability/maturity ladder, running adoption sprints, packaging learning into reusable artifacts, and making autonomy safe via sandboxing. developed
developed byAI as an Enterprise Operating System · p2
Frontier model capability now moves U.S. policy in weeks, not years (six weeks from light-touch policy to FDA-style review proposals after Mythos). developed
developed byHow the 4 New AI Models Change How You Work · p1
Tacit, undocumented organizational knowledge is treated as a strategic moat, and exposing it to foundation models is therefore an unrecognized capital expense developed
developed byWhy AI agents need a Context Lake · p2
Early diffusion-model deployments (Inception Labs' Mercury and Mercury 2, Google's Gemini Diffusion, open-source LLaDA) report roughly 1,000+ tokens per second on frontier GPUs, an order of magnitude faster than autoregressive single-stream decode. developed
developed byDiffusionGemma: 4x faster text generation · p3
Agent context should be treated the way modern software already treats its dependencies: versioned, pinned, and checked before it ships. developed
In large, complex tech organizations where developers already wrote comparatively little code and most complexity lived in organizational/system navigation, AI is unlikely to make a comparably large impact, unlike in simple, well-tested, low-risk codebases (the author's own 'a-ha moment' case). developed
Boris Cherny (creator of Claude Code) reports that in one month, 100% of his code contributions were AI-written — effectively fulfilling Dario Amodei's March 2025 prediction that AI would write 'essentially all code' within roughly a year, faster than most (including the author) expected. developed
By default, AI coding agents skip all of the invisible senior-engineer steps, taking the shortest path to 'done': asked for a feature, they write the feature, declare victory, and move on without a spec or a failing test first. developed
A verification budget should sequence checks by cost and value: fast, deterministic checks (linting, type checking) run early and continuously, while expensive but higher-value checks (full test suites, mutation testing, browser testing, security scanning) are reserved for later gates such as draft-PR time. developed
developed byBrownfield Agentic Engineering · p15
A human's own cognitive bandwidth does not scale with the number of parallel agent sessions they run, so parallel agentic work creates comprehension/cognitive debt and context-switching costs -- mental models going cold between sessions, decisions whose rationale isn't preserved anywhere -- that scale with session count, not with agent capability. developed
developed byBrownfield Agentic Engineering · p11
The recurring symptoms Osmani has named over recent months (comprehension debt, cognitive surrender, orchestration tax) are retroactively unified as instances of intent-debt management. developed
developed byOwn the Outer Loop · p1
Intent is structurally the one input that has to come from a human — agents can only fabricate plausible-sounding intent from the code, not generate the actual intent. developed
developed byAgentic Code Review · p8
Comprehension debt (the growing gap between what exists and what the human actually understands) grows faster, not slower, as a smooth loop ships more code the human didn't write, unless the human keeps reading what the loop produced. developed
developed bySoftware Factories, Light and Dark · p7
Two engineers can build the structurally identical loop and get opposite outcomes depending on whether they use it to move faster on work they understand deeply or to avoid understanding the work at all — the loop itself cannot tell the difference. developed
developed byWhat the Hell Is a Loop, Anyway? · p10
A loop requires five building blocks — Automations (scheduled discovery/triage), Worktrees (parallel isolation), Skills (externalized project knowledge), Plugins/connectors (tool access via MCP), and Sub-agents (idea/verify split) — plus a sixth non-negotiable: state/memory persisted outside the conversation, because the model forgets everything between runs but the repo doesn't. developed
Loop engineering sits "one floor above" agent-harness engineering: the harness is the environment a single agent runs inside, the loop is the outer system that runs the harness on a timer, spawns little helpers, and feeds itself. developed
developed byWhat the Hell Is a Loop, Anyway? · p1
Spawning agents is not the skill; designing the surrounding system around the single non-cloneable serial resource (your attention) is. developed
developed byThe Orchestrator's Tax · p8
Exposure 1 — upstream opacity: because third-party models are probabilistic, context-sensitive and drift while vendors rarely supply usable documentation of training data, failure modes and updates, the deployer inherits an unmanageable black box; the remedy is to build transparency requirements (intended use, training-data categories, limitations, test results, failure modes, material updates, compatible logs) into every AI contract and treat a vendor's inability to provide them as a procurement/governance risk, not a contracting inconvenience. developed
developed byIs Your AI Training the Competition? · p4
Load-bearing premise: liability tracks proximity to the end user, not causation within the technical stack — the enterprise closest to the affected person is treated as the accountable party regardless of which layer actually produced the harm. developed
A set of linked markdown files already forms a graph that an agent can traverse simply by reading and following links, making a dedicated graph database redundant for most cases. developed
developed byAdrien Taravant's Post · p5
The DORA / SPACE / DX Core 4 framework stack is a sufficient evaluation apparatus for internal platforms. developed
Composable self-service capabilities ('golden bricks') outperform prescribed end-to-end paths ('golden paths') for enabling fast flow. developed
The collapse in the marginal cost and skill required to build software is what makes previously not-worth-pursuing ideas (a joke to-do app, a personal CRM) newly viable to build at all. developed
developed byThe case for making your own apps · p1
A bloated AGENTS.md degrades coding-agent performance rather than improving it. developed
developed byAudit your Agent files · p1
An adversarial interview process (Grill Me) produces better shared understanding than the default plan mode in AI coding tools. developed
TDD forces LLMs to take small, disciplined steps that prevent entropy. developed
The report's own data reveals a leadership execution gap: leaders identify code maintainability, AI-output quality, and the junior pipeline as top concerns, yet organizational responses in both measurement (only ~a fifth using rigorous AI-impact metrics) and talent strategy (4% acting on junior-developer risk) remain shallow relative to the stated concern. developed
37% of leaders report increased hands-on technical work (up from 25% in 2025); 52% of advanced engineers (staff/principal/distinguished) report increased hands-on work (up from 30%); even managers-of-managers show a rise (25%, up from 18%). developed
Running multiple (3-10+) AI coding agents in parallel per person, with no token budget or usage tracking to promote 'tokenmaxxing,' is now normal working practice at Anthropic, and comes with very high individual autonomy relative to Big Tech or most startups. developed
developed byHow Claude Code Is Built · p11
Building a rigorous, shared, DSL-like vocabulary to communicate with the LLM — ideally co-built with the LLM itself — is a more promising line of attack than either unchecked vibe coding or large upfront specifications, echoing domain-driven design's 'ubiquitous language.' developed
By May 2026 [article's timeline, likely meant 2025], all three major AI coding harnesses (Codex, Hermes, Claude Code) had shipped a native /goal command that reproduces the Ralph loop's behavior in a single command, collapsing what used to require custom DIY setup. developed
developed byWhat the Hell Is a Loop, Anyway? · p1
The 2026 FDE role is becoming indistinguishable from a solutions architect or consultant. developed
developed byThis Week in AI: Production Viability · p13
The value proposition of routing is framed almost entirely as cost reduction, not quality improvement — the goal is to spend less while keeping quality merely 'good enough'. developed
Rust's hardest learning curve is data-structure design, not syntax — the borrow checker forces tree-or-DAG shapes and rejects naive cyclic graphs. developed
False-positive rate from security LLMs is dominated by two factors: programming-language memory safety (C/C++ produce more false positives than Rust) and model hedging bias. developed
Removing the exfiltration channel is the one actionable leg of the lethal trifecta, because untrusted input and sensitive-data reach cannot be removed without making the agent useless, and the channel cannot be removed entirely either — only contained deterministically. developed
The quiet exfiltration requires no exploited CVE, stolen token, or compromised process — only an outbound request the agent was always allowed to make; the agent does exactly what it was permitted to do. developed
Embedding a company's design system into AI tooling (via MCP, Cursor rules, bundled components) produces dramatically higher-fidelity prototypes. developed
Only 41% of surveyed employees combine AI-relevance with low psychological debt — suggesting majority of workers need combined business-case and debt-mitigation training developed
developed byLenny Rachitsky's Post · p3
Introducing cognitive friction (e.g., requiring hypothesis formation before AI use) preserves higher-order cognition and reduces cognitive debt developed
The team you build is the company you build — talent is the single most important variable in startup success. developed
Most companies that hire aggressively after raising money accomplish less, not more, because they add ammunition without adding barrels. developed
As AI becomes embedded in products, services, and decision-making, trust is emerging as a scarce, critical competitive advantage; companies that fail to earn it face regulatory action, reputational damage, and defection to more trustworthy competitors. developed
Local productivity (screens appearing, branches multiplying) can mask systemic degradation of project focus and maintainability. developed
Either governments will create AI governance, or a Three Mile Island-type disaster will force it developed
developed byThe Trump administration's AI doomer moment · p8
Lines of code is not a valid measure of AI-era productivity: AI's verbosity means more code measures 'slop,' not value, and all else equal, more code is more expensive to maintain and more bug-prone. developed
developed byMeasuring AI's Impact on Delivery Speed · p2
Pull-request count is not a valid measure of AI-era productivity, because its former loose correlation with development pace depended on engineers producing PRs at a human, 'consistent and logical' rate — a regularity AI breaks by producing more code than engineers can review. developed
developed byMeasuring AI's Impact on Delivery Speed · p2
Delivering software faster is not the same as delivering value faster: doubling the bug-fix rate does not have the same business value as doubling the rate of competitive new features, and speed applied to poorly-conceived features can make the product worse. developed
developed byMeasuring AI's Impact on Delivery Speed · p1
If AI simultaneously increases feature delivery and maintenance burden, a team's overall long-run productivity can spike and then fall below its pre-AI baseline, even though the short-term effect looks like a clear win. developed
developed byMeasuring AI's Unintended Consequences · p10
Burnout is a live risk of intensive AI use (e.g. under spec-driven development); burned-out engineers produce worse work, don't push back or report problems, and eventually leave, taking institutional knowledge with them. developed
developed byMeasuring AI's Unintended Consequences · p14
Rigorous measurement of AI's ROI is practically achievable by a typical engineering organization, not merely desirable in principle. developed
The future direction of AI pricing is genuinely uncertain — it could rise as subsidies end, or fall due to open-weight models, local inference, cheaper inference chips, routing, and more token-efficient models — so ROI models should state cost assumptions explicitly and test both optimistic and pessimistic scenarios. developed
developed byMeasuring AI's Unintended Consequences · p13
The 'AI jobs apocalypse' framing is wrong; industrialisation will create new industries via adjacent possibles (componentisation), not collapse the labour market. developed
The real problem with technology industrialisation in the West is inequality and absence of redistributive mechanisms — not the technology itself. developed
developed byThe Tidy House · p4
Codifying agent workflows into reusable, installable, versioned skills is a better unit for distributing engineering practice than ad-hoc prompting — the load-bearing premise behind shipping a versioned skills library at all. developed
Planning should use interactive intent elicitation rather than the human writing the plan: instead of plan mode, an agent grills the human one question at a time, explicitly separating facts (which the agent discovers by exploring the code) from decisions (which only the human can make), captures rationale in ADRs and a glossary, and gates on explicit human confirmation before implementing. developed
developed byHow I write software with LLMs · p7
Because of cost, agent-driven testing is currently better suited to targeted debugging and exploratory testing than to frequent execution in continuous integration pipelines. developed
Even the best harness will sometimes hit hard failures with small models, requiring cloud escalation as a fallback. developed
The harness, not the model, is the dominant lever in agentic coding performance — at least at the small-model end of the spectrum. developed
developed byExperiences with local models for coding · p2
Self-hostable, locally-run coding agents are a viable adoption path for serious software work, not just a hobbyist demo. developed
developed byExperiences with local models for coding · p3
Coding tasks are bounded enough that a 7B-20B model with a sufficiently constrained harness can complete them — i.e. the relevant agentic 'reasoning' is mostly about navigation, format compliance, and decomposition, not about latent intelligence. developed
Traditional junior developer training is breaking down because the simple entry-level tasks are disappearing developed
Human-legible naming and Ubiquitous Language likely still matter functionally in agentic development, not merely for human comfort, because language must be shared across code, specifications, diagrams, UIs, APIs and MCPs and across human/agent conversation — and because LLMs are trained on massive amounts of human language, not on hypothetical non-human-readable agent-only languages, which lack sufficient training data to be viable. developed
developed byWriting vs. Thinking · p11
Technical debt is not identical with bad or ugly code; code only becomes debt once you have to pay interest on it by repeatedly reading or modifying it, so interest rate is a function of time and requires a time dimension on the code to be measured at all. developed
Explaining an idea to the team before it becomes code is itself a forcing function for understanding — the Feynman 'freshman level' test: if you can't reduce your idea to something a Navigator or Driver can act on, you don't yet understand it as well as you think. developed
Software Teaming is not a panacea: it requires an already reasonably disciplined Agile team (TDD, frequent delivery, inspect-and-adapt), a non-punitive environment for experimentation, and genuine openness to differing ideas — absent those preconditions the authors expect it to fail, not merely underperform. developed
AI introduces a new kind of queueing — waiting on the AI/agent to finish — which teams should aim to eliminate, not merely learn to tolerate, the same way mob programming always treated 'what do we do while we wait' as a problem to design away rather than accept. developed
Whether code is now an ephemeral byproduct the agent regenerates from the spec, or remains the artifact that matters, is genuinely unresolved — the strongest counter to spec-as-center being that code, unlike a fuzzy prose spec re-interpreted each run, is unambiguous, executable, and deterministic, making the code the better spec. developed
As agents take the implementer role, the entry-level rung that trained junior engineers is disappearing, threatening the future supply of senior engineers. developed
A clear ambition about what an organization wants to be and where it wants to go is a foundational, timeless element of strategy that will never go out of fashion. developed
Strategy processes go badly when they become a purely analytical, dutiful exercise disconnected from people, and when groupthink causes a group to unconsciously converge on the same answer because it is too homogeneous or too averse to friction. developed
developed byStrategy Making Gets an AI Boost · p3
Longer working lives (50-60 years) combined with AI-accelerated change mean careers are becoming both longer and harder to predict, and leaders should respond with three practices rather than a single fixed plan: prepare before you need to (small, low-stakes experiments), commit to what lasts (durable capabilities and relationships), and protect time for judgment (deliberately protected reflective 'calm'). developed
developed byLiving and Working in the Age of Longevity · p12
Employees under structured empowerment should be assessed on key outcome results — whether they delivered customer value — not on which options they chose or process compliance. developed
developed byPim de Morree's Post · p5
Focal employees should be involved in identifying the options on the menu and periodically invited to provide feedback, creating a continuous cycle of local experimentation and organizational adaptation. developed
Attention is largely volitional — the barrier is a matter of choice/will, not aptitude or circumstance. developed
developed byUten kontroll · p6
Gut judgment that contradicts a rationalized head decision is a reliable signal that should usually be heeded developed
The decisive work of engineering leadership under AI — judgment about what to build, sound operating principles, and engineering rigor — is unchanged in essence; AI shifts the magnitude, economics, and tooling, not the fundamentals. developed
developed byRevised rules of engineering leadership. · p9
When teams remove friction, seek quality, and want to deliver value quickly, TDD, TBD, and team-focused development emerge as natural convergent practices. developed
Legitimacy is categorically different from legibility — it asks whether control is justified, not whether it is visible developed
AI will not fix the funnel — it reveals and amplifies whatever version of it you are already running developed
Understanding the return on engineering investment is a set of continuously practiced behaviors, not a calculation you perform. developed
Organizations that invest in codebase maintenance ('gardening') compound their advantage and avoid the entropy that cripples peers. developed
A bad proxy (like story points as investment theory) is worse than no proxy at all because it provides false confidence in a wrong answer. developed
The appropriate epistemic framework for engineering ROI is Bayesian updating -- starting with priors, collecting signals, narrowing uncertainty quarterly -- not demanding definitive proof. developed
AI is best thought of (per Kent Beck, cited approvingly) as an imperfect, eager-to-please genie — useful but not autonomous, requiring direction and verification. developed
developed byTBM 428: Yes, Robot. Yes, Boss? · p4
Solo AI work is deep but narrow — AI extends your cognition powerfully but holds only what you gave it, not what three people with different expertise would have given it. developed
developed byTBM 439: Day At The Gig... · p6
Nobody currently has a valid AI playbook, regardless of how confidently they market one. developed
developed byJohn Cutler's Post · p7
Most organizational systems (org charts, taxonomies, planning tools) are structurally incapable of holding more than one frame at a time; they flatten multi-frame reality into a single canonical view. developed
developed byTBM 434: How Maps Can Hide Problems · p3
Honesty in organizational description is a moral/political act, not an analytical one — and the analytical framings on offer (capability maps, journey maps, DDD) become traps when used to substitute for that act. developed
developed byTBM 434: How Maps Can Hide Problems · p6
The gap between aspirational and actual org models may be functional, not pathological — Brunsson's 'organized hypocrisy' lets contradictory stakeholders be satisfied simultaneously. developed
Organizational change is gated by relational and political conditions — safety, alignment, power, timing, stewardship, feeling seen — not by information availability; so improving visibility/legibility does not, by itself, move the system. developed
In a collective there is no single mirror and no 'view from nowhere' — there are dozens of people holding up partial, angled fragments, each reflecting something real but partial, and each with something at stake in what gets seen. developed
Whether AI use in a given system is beneficial or harmful cannot be assessed from the technology alone; it requires first diagnosing which of the four quadrant/trap positions the asset already occupies. developed
developed byTBM 431: The Denominator That Matters · p11
Liquidation is not inherently a failure; refusing to liquidate -- continuing to build an asset that has never proven its value (the Build Trap) -- is the more common and more damaging failure. developed
developed byJohn Cutler's Post · p14
Portfolio health is not just a snapshot of the current mix of assets -- it is also a function of what is flowing between categories, how fast, and whether anyone is intentionally managing that movement. developed
AI changes the magnitude, cost, and feasibility of each capability motion, but does not change the underlying physics -- the causal structure -- of the pattern itself; the appropriate question is motion-by-motion ('to what extent, under what conditions, with what second- and third-order effects'), not a single yes/no verdict. developed
developed byTBM 431: The Denominator That Matters · p1
Less experienced people (Believers) accept the performed strategy not out of naivety about organizational politics specifically, but because they lack the tacit skill to recognize what a real strategy looks like -- if it sounds coherent enough, they take it at face value. developed
When strategy, structure, technology, and incentives are coherent, an organization can leave a great deal implicit — one map (e.g. team or goal) reliably lets you infer information that technically belongs to a different map. developed
These coping mechanisms can lower the pain of incoherence enough that the underlying incoherence never actually gets addressed — coping substitutes for repair. developed
A defensible AI business case is not a single point-ROI estimate ('$10M spent -> $30M value') but a set of explicit, falsifiable 'theory of the system' hypotheses, from which ROI is properly the observed output rather than the initial pitch. developed
developed byJohn Cutler's Post · p2
People systematically observe adaptation (layer 2), attribute it to belief (layer 3), and underestimate the causal role of power-weighted norms (layer 1). developed
Software subsystem boundaries ('fracture planes') should be chosen to match team-first boundaries — what a team can hold at acceptable cognitive load — rather than purely technical decomposition logic. developed
The four-topology taxonomy implicitly assumes every team's members are human, with cognitive load arising from a human mind processing the task — it has no category for a team whose 'members' include autonomous AI agents with a different cognitive-load profile (no fatigue, but no implicit domain context either). developed
Cognitive load for software work decomposes into intrinsic, extraneous, and germane load; extraneous load should be minimized (it is pure waste) while germane load should be protected (it is where learning and good decisions happen). developed
There are exactly four fundamental team topologies — stream-aligned, enabling, complicated-subsystem, and platform — and most other named team types in practice are variations or combinations of these four. developed
developed byDevOps Topologies · p2
A platform should aim to be the 'Thinnest Viable Platform' — the smallest amount of shared, self-service capability that actually reduces cognitive load for consuming teams — because over-building the platform creates a new bottleneck. developed
developed byDevOps Topologies · p4
Because current AI tools can act directly on a person's files (reading, renaming, moving, cross-linking) rather than only answering chat questions, they amplify whatever organizational structure already exists in those files. developed
Once governance is externalized into deterministic layers, model intelligence stops being a prerequisite for safety and becomes a pure, tunable economic variable, formalized as Total Decision Cost = Compute Cost + (Escalation Rate × Human Cost). developed
Staff engineers are the most effective agent supervisors but are underutilized because they spend disproportionate time on human coordination developed
AI-enabled productivity is causing a regression toward larger batch sizes and waterfall-like patterns, reversing a decade of DORA findings developed
Conway's Law applies to agents — organizational design must now account for agent mobility, specialization, and drift developed
A schema captures only the shape of a tool's arguments, not its meaning; it cannot tell an agent when a tool is appropriate, how its effects interact with other tools, or whether two similarly named operations are semantically equivalent — portability across hosts does not guarantee consistent behavior across models and contexts. developed
developed byMCP Is Not Just Another API Standard · p6
When work and life demands conflict simultaneously there is often no objectively 'right' choice, and the mental energy spent on that trade-off itself drains capacity independent of the work actually performed. developed
developed byDare to Be Bad at Something (From Fixable) · p13
Provisional employment (internships, co-ops, contract-to-hire) is currently the best available solution to the hiring-signal problem, because sustained real work generates dramatically more reliable signal than any interview loop. developed
developed byThe Case for Hiring Job Hoppers · p1
Because every run can trace a different path through the same prompt, fixed call graphs and clean service maps assume a regularity agent behavior doesn't have, so request-level monitoring can show ten thousand successful calls and never notice the one loop that ran far longer than it should have. developed
A useful observability substrate needs a minimum concrete signal set -- run/turn structure, token and model accounting, tool-call events, guardrail decision events, identity and authority context, and outcome/bound metadata -- landed in a fleet-queryable store, not only a per-run trace viewer. developed
AI is not an 'AI strategy' or 'tool selection' problem at L1→L2 — the blocker is that the company cannot describe its own workflows developed
Flat-rate and per-seat AI pricing models are unsustainable and will be replaced by usage-based billing developed
developed byHow to Respond to the Coming AI Cost Shock · p2
Retry storms — where individually rational client behavior produces collectively catastrophic outcomes — can only be prevented by enforcing backoff and circuit-breaking as SDK defaults, not by documenting the correct strategy. developed
Organizations that invested in clean data infrastructure before LLMs can apply AI immediately; those that didn't pay GPU costs reconstructing context they should already have. developed
developed byAI as an Enterprise Operating System · p8
Decisions to persist, pause, or pivot on a commitment should be made through a repeatable metacognitive review process (Plus Minus Next) rather than by defaulting to consistency with past commitments. developed
Authorization for AI agents must happen at runtime — evaluating each call as it occurs — rather than at session start. developed
developed byThe Meter Was Always Running · p7
Because compounding correctness flips the calculus, maximal token spend becomes ROI-rational, so tokenmaxxing will return long-term — this time driven by genuine value rather than incentive-gaming. developed
Shared understanding should be treated as a first-class deliverable of software development -- something teams invest in explicitly, with tooling, time allocation, and design principles -- rather than something that happens as a side effect of writing code. developed
Generative AI is simultaneously reducing technical debt (via automated refactoring, AI-assisted review, and test generation) while accelerating the accumulation of cognitive and intent debt, because AI generates code faster than teams build genuine understanding of it or externalize the rationale behind decisions. developed
developed byWho cleans up after the vibe-coding party? · p9
Anthropic's refusal to grant blanket 'all lawful use' terms to the Pentagon was principled rather than obstructionist, as the very capabilities it sought to gate are now recognized as genuinely dangerous. developed
developed byWalter Isaacson finally got scared about AI · p6
The platform/runtime is now a more decisive engineering choice than the LLM itself — the seven-day persistence and sub-second cold-start claims depend on infrastructure, not the model. developed
The author frames the founding motivation autobiographically: having helped build the RLHF-era methods behind ChatGPT at OpenAI, the author concluded that chat-optimized LLMs were missing something essential for automating real-world software decisions, motivating a four-year effort (two years in stealth) culminating in Jev. developed
The author became clinically addicted to social media, meeting 'frighteningly many' of the standard signs of addiction — compulsion, inability to control amount/time/stopping, withdrawal, tolerance, sacrificing other interests, and continued use despite recognised negative consequences. developed
Attention and focus are a depletable, hijackable resource: sustained capture measurably degraded her concentration, focus and memory and crowded out reading, hobbies, relationships and sleep — real costs that persist even when clinical mental-health measures (e.g. depression) register nothing. developed
Making step checkpoints a primary-key/unique constraint lets the database enforce idempotency, so a step that reruns after a crash reads its prior result instead of repeating a side effect. developed
Crash recovery can be implemented with a lease-and-sweeper pattern: workers heartbeat the rows they own, and a periodic query re-enqueues any execution whose lease expired. developed
The inference/validation/query/integration payoff is real and worth the ontology overhead. developed
A vault can progressively enrich (prose -> typed fields -> linked meaning -> graph projection) and, exposed as DPROD data products, scale unchanged from a private second brain to team memory to enterprise integration to an open decentralised agentic web. developed
SBOM signing is more important than the specific signing tooling used — any signing is better than no signing. developed
The gap between available verification tooling and actual adoption is the primary problem — the tooling exists, the practice does not. developed
vlt's phased-install design assumes that automatic execution of lifecycle/install scripts is the primary mechanism by which supply-chain malware compromises developer machines and CI systems. developed
The task loop (Geoffrey Huntley's 'Ralph loop') restarts a coding agent against the same spec with a completely fresh context window every iteration, and the apparent waste of refeeding the full spec each time is deliberate — it prevents the context rot and compaction events that degrade long-running sessions. developed
The developer whose job was writing boilerplate or scaffolding CRUD apps is permanently gone; AI models can now do that work in seconds. developed
Architects and developers are converging into a new role, 'orchestrator,' in which one person directs work that previously required entire teams. developed
The behavior does not have to change for the verdict to change: extraordinary results upgrade harsh behavior into excellence or even legend, while a downturn reclassifies the identical behavior as misconduct. developed
developed byEvery Organization Has Two Realities · p5
Effective leadership requires a form of therapeutic self-awareness analogous to what clinicians develop regarding countertransference. developed
ASR performance scales strongly and roughly monotonically with model size for Norwegian (Tiny through Large), for both OpenAI Whisper and NB-Whisper, but the absolute and relative gap between the two narrows as size increases. developed
An individual enterprise engineer using coding agents is worth roughly 100x a consumer subscriber in token economics. developed
This dynamic is a tragedy of the commons: individual developers capture personal productivity gains from AI-assisted contribution while externalizing the resulting review, verification and maintenance costs onto reviewers, maintainers and the wider community. developed
Some maintainers (tldraw's Steve Ruiz) have responded by banning outside AI-assisted contributions entirely, arguing that once writing code is the easy part, there is no reason to want someone else -- or something else -- to write it, and that the social contract underlying open-source contribution is 'obliterated.' developed
The dark-factory debate reduces to whether agents are more like compilers (regenerating code from scratch when needed) or more like engineers (maintaining and moulding an existing codebase over time). developed
Your success in life is determined largely by your ability to speak, your ability to write, and the quality of your ideas — in that order. developed
developed byDo you even need a presentation? · p8
Blackboards are the right tool for informing/teaching; slides are for exposing ideas (job talks, conferences). developed
developed byDo you even need a presentation? · p6
Automatic updates, while beneficial for patching known vulnerabilities, create a force-multiplier for supply-chain attacks by ensuring compromised code propagates to all users without review. developed
Grok Skills is more like a reusable workflow/capability layer than a fully deployable autonomous agent system developed
Historically, whenever the cost of producing information collapses (Gutenberg's press, cheap paper, spreading literacy), the prior scarcity-based gate stops working and society invents new socio-institutional gates — exemplified by the curated journal (Oldenburg's Philosophical Transactions, de Sallo's Journal des sçavans). developed
AI did not break a working gate — the popularity-proxy gate was already dysfunctional — but AI is now flooding every domain that relied on scarcer, human-judgment-based gates (academic journals, open-source repos) simultaneously with the already-strained algorithmic platforms. developed
Leaning harder on existing human-centric institutional gates (reputation, credentials, follower counts, h-index, personal networks) is the most likely near-term response, but it entrenches incumbents and undermines the democratization of contribution that cheap AI-assisted production made possible. developed
developed byWhen Content Is Free, Trust Is the Product · p3
An AI gatekeeper also creates an adversarial incentive problem: once both the generating side and the gate-checking side use powerful AI, each side is incentivized to build systems that deceive the other (a deceptive job-applicant AI vs. a hiring AI; a p-hacking-friendly research AI vs. an editorial-review AI). developed
developed byCo-Existence and the End of Co-Intelligence · p7
Abbott's core claim: professions compete for 'jurisdiction' — the right to own certain kinds of work — and this contest is fought in three arenas: the law (licensing/bar rules), public opinion (whether people believe they need a lawyer), and the daily work (who actually performs the task). developed