Fragments: May 5
Martin Fowler · rss · 2026-05-05 source ↗ Reader 🖥️
Agent SkillsCompute & Energy InfrastructureFeedback LoopsThe Genie (Beck)Local AIMalleable SoftwareSandboxing
The current moment in AI-assisted software development is defined by a tension between the power of cloud-hosted AI genies and the risks they carry — to code quality, to truth, to budgets, and to autonomy — and the pieces Fowler curates collectively ask whether the escape route runs through disciplined engineering practice, local-first computing, or some combination of both.
theoretical · link-commentary digest / curated fragments — A weekly roundup where Fowler selects, quotes, and contextualizes work by others — adding editorial commentary that builds a coherent theme across seemingly unrelated links. The genre is closer to a curated newsletter than an essay, but the editorial voice imposes argumentative structure.
Structure
| Part | Role |
|---|---|
| Lattice framework (Rahul Garg / Reduce Friction in AI) | Opens with a concrete engineering response to AI tool deficiencies: composable skills that embed standards and accumulate project context |
| SPDD Q&A addendum | Brief note on structured prompting generating enormous interest and follow-up questions |
| Jessitron's double feedback loop | Identifies the meta-level loop: AI tools let developers reshape the tool itself, not just the product |
| Ashley MacIsaac / Google AI defamation | Pivots to the harm side: AI-generated content fabricated defamatory claims about a real person |
| RedMonk AI infrastructure spend | Quantifies the staggering scale of cloud AI investment (50-75% of revenue) and notes Apple as outlier |
| Willem van den Ende: local models are Good Enough | Argues that open local models have crossed the usability threshold for daily coding work |
| Apple's hardware strategy / Nate B Jones | Reframes Apple's low AI spend as a deliberate replay of the Apple II vs. mainframe strategy |
| Kent Beck's Genie Tarpit | Closes with the deepest concern: AI tools produce code that lacks internal quality, leading to a complexity tarpit |
Key terms
- Internal Reprogrammability Malleable Software
The ability to reshape one's own development environment from within — a property of Smalltalk/Lisp cultures that agents are reviving - double feedback loop Feedback Loops
Two concurrent loops: (1) building/checking the product, (2) reshaping the tool/process itself when friction is felt - Genie Tarpit The Genie (Beck)
Beck's metaphor (via Brooks) for how AI assistants produce plausible code that accumulates hidden complexity until even the AI cannot make progress - Lattice Agent Skills
Garg's open-source framework of composable skills (atoms, molecules, refiners) that encode engineering standards and accumulate project-specific context for AI assistants - Zero Trust Architecture (applied to AI tools) Sandboxing
Sandboxing AI coding agents (e.g., Nono) on the assumption that powerful tools should not be implicitly trusted
Propositions
- explicitAI coding assistants silently make design decisions, forget constraints, and skip review — but composable skill frameworks like Lattice can embed engineering discipline that accumulates with use.conf: high
- explicitAgentic programming is recovering the Great Lost Joy of molding one's own development environment — the Internal Reprogrammability that was central to Smalltalk/Lisp but lost in the IDE era.conf: high
- explicitWhen AI tools generate and publish factual claims about real people, the company operating the AI bears defamation responsibility — it is not merely a neutral search engine.conf: high
- explicitBig tech companies are spending 50-75% of revenues on AI infrastructure, but Apple is a dramatic outlier at roughly 10%, which may reflect a deliberate bet on local-first AI.conf: high
- explicitLocal open-source models have crossed the 'good enough' threshold for daily coding work, making the cloud-AI dependency optional rather than necessary.conf: medium
- explicitApple's low AI infrastructure spend may be a replay of the Apple II strategy: putting less-capable compute locally while others invest in centralized mainframe-equivalents.conf: medium
- explicitAI assistants naturally produce code that sits below the quality threshold needed for sustainable software — complexity accumulates until even the AI cannot pretend to make progress.conf: highsupportsIt Ain't Broke: Why Software Fundamentals Matter More Than Ever · p2 — Both claim unguided AI code accumulates complexity/entropy until progress stalls.
- explicitdevelopedIt 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.conf: highrefinesIt Ain't Broke: Why Software Fundamentals Matter More Than Ever · p1 — Fowler reopens as an explicitly unresolved question what Pocock asserts flatly: whether internal code quality still matters in agentic programming.extends ←Does code quality still matter? · p4 — Fowler leaves open whether internal code quality still matters in the agentic age; Seemann extends the question to its logical extreme — a future with no human coders at all — and argues the human-cognition-based notion of quality specifically would become irrelevant there, while a structural analogue might not.
Arguments
The discipline-and-structure approach to AI-assisted development (Lattice, SPDD, double-loop tooling) is necessary because AI tools left unconstrained produce unsustainable code. strength: medium
Explicit premises:
- AI assistants skip design, forget constraints, and produce code nobody reviewed against real standards (Lattice framing).
- Structured prompting (SPDD) has generated enormous interest, indicating broad professional demand for such discipline.
- The double feedback loop means developers can reshape their AI tools to enforce standards and reduce friction.
- Beck observes that AI tools naturally produce code below the quality threshold — the Genie Tarpit.
Implicit premises:
- Internal code quality has historically been necessary for sustainable software development.
- Frameworks and structured approaches can compensate for what AI assistants lack in engineering judgment.
Inferences:
- If AI tools tend toward the tarpit and discipline frameworks can counteract this tendency, then adopting such frameworks is a rational response to the current state of AI-assisted development.
The future of AI may be local and hardware-centric rather than cloud-based, and Apple is strategically positioned for this outcome. strength: medium
Explicit premises:
- Big tech is spending 50-75% of revenue on cloud AI infrastructure; Apple spends ~10%.
- Local open-source models are already 'good enough' for daily coding agent work.
- Apple's new CEO (John Ternus) is a hardware leader, signaling a bet on sophisticated local hardware.
- The Apple II precedent shows that less-capable local compute can displace centralized systems by enabling entirely new use patterns.
Implicit premises:
- The cloud-AI infrastructure spend is not sustainable indefinitely at 50-75% of revenue.
- Privacy and data sovereignty concerns favor local computation.
- Open-source model quality will continue to improve.
Inferences:
- If local models are already good enough and improving, while cloud spend is unsustainably high, the market may shift toward local AI — and Apple's hardware focus and low cloud spend would then be prescient rather than negligent.
Narratives
- The Apple II vs. mainframe replay — Apple's CEO transition and hardware focus as a deliberate bet that AI will follow the same path as personal computing: from centralized to local.
→ Transforms Apple's apparent AI underinvestment from a weakness into a strategic masterstroke, using historical pattern-matching to lend it credibility. - Ashley MacIsaac's defamation by Google AI — a real musician falsely labeled a sex offender, losing a concert and fearing for his safety on stage.
→ Makes the abstract risk of AI-generated misinformation visceral and personal, grounding the accountability argument in genuine human harm. - Brooks's tar pit, summoned by Beck — great beasts struggling against the tar, fiercer struggle producing deeper entanglement.
→ Connects the current AI coding moment to the oldest and most resonant metaphor in software engineering, implying the problem is not new — only the beasts have changed.
Questions addressed
- ✅ How can engineering discipline be embedded into AI-assisted development workflows?
Composable skill frameworks (Lattice) and structured prompting (SPDD) encode standards that accumulate with use; the double feedback loop lets developers continuously reshape their AI tools. - ✅ Who bears responsibility when AI generates and publishes false claims about real people?
The company operating the AI (Google, in the MacIsaac case) must take responsibility — it is publishing, not merely searching. - ⭕ Is the massive cloud AI infrastructure spend by big tech sustainable?
Flagged as a tension: Apple's outlier position and the local-model thesis suggest the current spend pattern may not be the equilibrium. - ⭕ Does internal code quality still matter in the age of agentic programming?
Explicitly left open. Fowler presents both sides: Laura Tacho's view that DX and agent-experience overlap, versus the galaxy-brain view that LLMs will transcend spaghetti code. - ⭕ Can AI coding agents escape the complexity tarpit, or will they sink like every other tool before them?
Posed as the 'fundamental question' in the closing paragraph. Left deliberately unresolved.
Gaps
- unstated assumptionFowler's curation implicitly assumes that the discipline-vs-tarpit tension is the defining axis of AI-assisted development, but economic and organizational dimensions (team structure, hiring, skill deskilling) are largely absent from the fragments selected.
Read next: Cross-reference with pieces on how AI changes team composition and organizational design (e.g., the 'kill your org chart' literature). - implicit premiseThe local-model argument rests on a single practitioner's 'good enough' assessment (van den Ende). There is no quantitative comparison of local vs. cloud model performance on real engineering tasks, leaving the threshold claim unsubstantiated.
Read next: Look for benchmarks comparing local open-weight models (e.g., Llama 3, Mistral) against cloud models on SWE-bench or similar agentic coding evaluations. - tension not resolvedThe piece opens with frameworks that impose structure on AI output (Lattice, SPDD) but closes with the possibility that internal quality may not matter at all. These two positions are in direct tension, and Fowler does not reconcile them — the curatorial format lets the contradiction stand.
Read next: Track whether Fowler takes a stronger position in subsequent Fragments or dedicated articles; compare with Beck's evolving stance on AI and internal quality.closed byThree more static code analysis sensors — A subsequent martinfowler.com piece takes the position the Fragments left open: internal quality still matters because agents reliably compound modularity debt without enforced rules and periodic AI review. - selection biasThe fragments are curated from Fowler's professional network (ThoughtWorks colleagues, Beck, Jessitron, RedMonk) — a community that values craft and internal quality. The 'move fast and let AI clean up later' perspective is represented only as a foil, not through its strongest advocates.
Read next: Seek out arguments from the 'galaxy brain' camp — practitioners or companies that have successfully scaled with AI-generated code despite low internal quality.
Related documents
- It Ain't Broke: Why Software Fundamentals Matter More Than Ever
- Have your agent record video demos of its work with shot-scraper video
- Starving Genies
- Building AI Agents For The Enterprise
- Cybersecurity and the Gap Between Skill and Ability
- Experiences with local models for coding
- Presentation: Fixing the AI Infra Scale Problem by Stuffing 1M Sandboxes in a Single Server
- Three more static code analysis sensors
- Google Announces GKE Agent Sandbox and Hypercluster at Next '26
- How Do You Know That?