The AI paradox: More automation, more humans, more work | Dan Shipper

Lenny Rachitsky (host); Dan Shipper (guest, CEO Every) · podcast · 2026-05 Reader 🖥️

Agent-to-Agent Protocols AI Labor Disruption AI Slop Code Was Never the Bottleneck Evals Forward-Deployed Engineer Human as Orchestrator of Agents Near-Zero-Cost Code Product Engineer (PM/Engineering Role Merge)Riding Model Improvements

Despite radically improving model capabilities, the agentic era will paradoxically generate more work for more humans — because every agent needs a human gardener, SaaS persists (with users bringing their own tokens via Codex/Claude Code), and progress merely commoditizes yesterday's competence and pushes humans further up the value chain to define what is worth doing next.

practical · mixed — A predictive conversation framed as Shipper's year-ahead forecasts (three buckets: how work will be done, the shape of work, who will succeed). Its rhetorical thrust is prescriptive ('ride the models', 'try Codex', 'be a generalist', 'PMs and full-stack designers win'); its evidentiary backbone is anecdotal — Every's internal practices serve as the case study from which all generalizations are projected. Mixed expository/argumentative form within a dialogic frame.

Structure

Part	Role
Opening framing: Shipper's previous correct call (Claude Code for non-engineering work) and Every's setup (~30 AI-pilled employees, doubled in a year) as the lens for prediction	Establishes Shipper's predictive credibility and the empirical basis ('living in the future') for everything that follows
Bucket 1 — How we will work: bifurcation into (a) a per-company super-agent in Slack and (b) Codex / Claude Code / Co-work as the desktop work surface, with SaaS rendered inside an agent's in-app browser	Core predictive content; redefines the substrate of knowledge work
Sub-argument: 'SaaS apocalypse is dumb' — agents increase SaaS demand, users bring their own tokens, margins are protected; counter-intuitive corollary that SaaS is a buy	Contrarian financial/strategic claim derived from the agent-as-surface prediction
The 'automation is a lie' / 'every agent needs a human' thesis, evidenced by Every's hiring doubling and the senior-engineer benchmark anecdote	Names the central paradox and provides its mechanism (the reach test, the benchmark goalposts, vibe-coded slop creates more work)
Bucket 2 — The shape of work: pull-request explosion from non-technical staff; forward-deployed engineer as new role; everyone-can-do-everything generalism creating role-identity confusion; AI-generated writing becoming normal and welcome	Maps second-order organizational effects
Bucket 3 — Who wins: PMs (Marcus story), full-stack designers, generalists; the 'ride the models' / 'turn the rock over' / 'find your moment of joy' prescriptions	Practical conclusion / call to action
The 'frozen human competence' theory: models commoditize yesterday's competence; humans always re-establish a frontier above it	Theoretical scaffolding for the optimistic labor-market view
Closing reflections: 'everything has changed and nothing has' / Middle Ages horizon analogy; lightning round (books, motto: 'spaciousness and strength')	Rhetorical close that anchors the both-and posture

Key terms

Ride the models Riding Model Improvements
Shipper's prescription for individual durability: continuously try each new frontier model on whatever you do, treat it as a new power to extend your reach, and 'turn the rock over' periodically because what failed last release may now work
Reach test
Shipper's informal evaluation criterion for whether a tool has crossed into real utility: when you wake up and reach for it organically without prompting
Super agent (one-agent-per-company) Human as Orchestrator of Agents
A single company-wide agent (e.g. Shopify's, Ramp's, Every's 'Claudie') maintained by a forward-deployed engineer, contrasted with the personal-agent model Shipper previously favored and has now abandoned because 'agents need a human who cares about them'
Co-work / Claude Code / Codex (as work surface)
Agentic IDE-style environments running on the user's machine with an embedded browser, becoming the 'operating system for work' — not just for code but for email, documents, research; the human and the agent share the same in-app browser
Forward-deployed engineer (FDE) Forward-Deployed Engineer
A specific human role responsible for keeping a company's super-agent working, debugging it, adding context, and translating user needs; the structural reason 'every agent needs a human' generates net new headcount
Automation is a lie Code Was Never the Bottleneck
The slogan-form of the paradox: every automation requires a human on top to make sure it works; benchmark progress mistakes 'autonomy' for genuine independence from human attention
Allocation economy Human as Orchestrator of Agents
Shipper's earlier coinage for the era in which humans relate to AI as managers do to employees — and crucially, managers actually work hard, they are not 'on the beach'
Frozen human competence Near-Zero-Cost Code
What models contain: the compressed, ingestible past output of humans. Each new model release makes 'yesterday's human competence cheap', which commoditizes it and forces humans to push to a new edge
Senior engineer benchmark Evals
Shipper's personal eval: ask a frontier coding model to rewrite his vibe-coded app Proof from first principles, score it against two human senior-engineer rewrites. Humans score high 80s/low 90s; pre-GPT-5.5 models scored ~30; GPT-5.5 with an Opus 4.7 plan scored 62 — and Shipper says when models 'saturate', he will redesign the benchmark and zero them out
AI-generated document / slop AI Slop
Distinction between AI-generated artifacts the author actually owns (read, edited, stands behind) and 'slop' (took less time to make than to read; sender cannot defend any line). The first is welcome and often better than human prose; the second is the new social violation
Edge of AI
Not San Francisco labs but anywhere a real human applies a new model to a real task; the discovery work of what a model can do happens at the user, not the maker
Two-agent collaboration Agent-to-Agent Protocols
The pattern where the user's agent (Codex) talks to a product's agent on the user's behalf, carrying compressed context the user could never type — predicted to become the default integration pattern rather than custom in-app agents
Spaciousness and strength
Rob Brezsny-derived posture for facing hard things (including AI's career disruption): not the David Goggins push-through, but a settled internal stance from which to engage

Propositions

explicitWork will bifurcate into two agentic surfaces: a per-company super-agent (most likely in Slack) and a Codex/Claude Code/Co-work desktop environment where humans and agents share a browser.conf: high
explicitThe personal-agent model fails because agents currently require a human who cares enough to maintain them; therefore the workable near-term pattern is one super-agent per company, maintained by a forward-deployed engineer.conf: high
explicitAutomation is a lie: every automated workflow generates net new human work above it (supervision, context-keeping, judgment calls, system design).conf: high
explicit⚡ disputed & corroboratedHiring at AI-native companies is going UP, not down — Every doubled headcount (15 -> ~30) in the past year despite being AI-forward.conf: high
supportsPresentation: Leadership in AI-Assisted Engineering · p7 — Every's doubled headcount at an AI-forward company is a live case of the claim that successful AI adopters should hire more engineers, not fewer.
supports ←AI Companies Are Hiring More · p7 — Cross-firm Ramp/Revelio data (high-AI-adoption firms grew headcount ~10%, entry-level ~12%) corroborates Shipper's single-company claim that AI-native hiring is going up.
contradicts ←AI Is Changing the Rules of Entrepreneurship · p5 — This piece's YC-study finding that AI-native-built startups run 25% smaller (fewer entry-level workers and managers) sits in tension with Dan Shipper's report that hiring at his AI-native company is going UP, not down (headcount doubled in a year) -- though the tension may resolve by growth stage: lean at YC-scale founding, hiring again once product-market fit and scale hit.
explicitThe SaaS apocalypse is wrong; agents increase the number of users of SaaS rather than replacing it, and users bringing their own tokens via Codex/Co-work actually protects SaaS margins.conf: high
explicitCLIs as a primary work surface are a transient era; GUIs (Codex/Co-work) will win for the same use cases, including for non-programmer work.conf: high
explicitcontestedBenchmarks 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.conf: high
contradicts ←Kaggle is making AI benchmark creation effortless · p5 — Kaggle claims more benchmarks drive progress in the areas that matter; Shipper argues benchmarks structurally miss the human framing work that matters most.
explicitModels commoditize yesterday's human competence; the structural consequence is that humans permanently move up to a new frontier rather than being displaced.conf: high
supportsThe keyboard was never the bottleneck · p3 — Same claim in different vocabulary: AI commoditizes the lower levels of work and pushes humans permanently up to the higher ones.
explicitcontestedProduct managers and full-stack designers are the role categories best positioned to gain leverage from agentic tooling (the Marcus / Spiral example as proof point).conf: high
contradictsHard truths about building in the AI era | Keith Rabois (Khosla Ventures) · p3 — Rabois says AI collapses the PM translation layer and eliminates the role; Shipper says PMs are among the biggest winners in leverage from agentic tooling.
explicitestablishedGeneralists with curiosity beat specialists in this transition, because the value moves to figuring out what to build and ensuring coherence.conf: medium
supports ←Head of Claude Code: What happens after coding is solved | Boris Cherny · p12 — Both claim that curious generalists outperform narrow specialists in the agentic-AI transition; this interview's advice to builders independently corroborates the other Lenny's Podcast episode's claim about who gains leverage.
refines ←Netflix CPTO on AI and the future of product and tech roles | Elizabeth Stone · p9 — Stone agrees generalists are ascendant but adds a reservation Shipper's episode doesn't carve out: a small set of genuinely irreplaceable deep specialists remain necessary.
supports ←Adam Mosseri: AI is a tailwind for authenticity · p10 — Both: curious generalists win the transition because value moves to figuring out what to build; Mosseri elevates curiosity/willingness-to-try as the top trending-up hiring trait.
explicitAI-generated writing inside an organization (plans, emails, strategy docs) will become normal and often better than human prose, provided the sender owns/edits and can defend every line.conf: medium
explicitSales (and to a lesser extent CEO/middle-manager) roles are the least changed by agentic AI so far; engineering, design, data science, PM are the most changed.conf: medium
explicitA CEO's company will go only as far as that CEO's personal AI fluency goes; AI fluency cannot be delegated.conf: high
explicitThe right unit of analysis when comparing 'AI vs humans' is always one-human-with-AI vs another-human-with-AI; pure AI-vs-pure-human comparisons are category errors because AI does not use itself.conf: high
explicitTwo agents (the user's and the product's) collaborating is structurally better than one agent embedded in a product, because the user's agent already carries compressed personal context.conf: medium
explicitWorkflow software must be redesigned for simultaneous human+agent collaboration (approvals, agent-readable HTML, CLI surface, change inboxes, rollback) — not as 'add an AI feature' to legacy UIs.conf: medium
explicitThe job apocalypse narrative is wrong and motivated; AI-CEO doom messaging is both ineffective and substantively mistaken.conf: high
implicitestablishedThe 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.conf: high
supports ←The Conductor Developer · p2 — Both identify human attention as the bottleneck on AI-assisted productivity that does not relax as the tools improve — Laycock frames it as the terminus of a sequential bottleneck migration through the SDLC, Shipper/Lenny frame it as a standing implicit premise of agentic productivity generally.
supports ←The Orchestration Tax is You · p2 — Osmani's judgement-throughput-as-binding-constraint is the same claim as Shipper's implicit premise that human attention is the non-relaxing bottleneck on agentic productivity.
implicitHeadcount-as-leverage-indicator persists in the agentic era — Shipper takes 'we doubled' as evidence of strength, not failure to automate.conf: high
implicitEvery's operating model — a media-and-software company of curious early-adopter generalists who get pre-release model access — generalises to most knowledge-work organizations.conf: medium
implicitToken-billing-to-the-end-user (Codex/Co-work pays for the SaaS interaction) will remain a stable economic equilibrium rather than collapse back into vendor-charged consumption.conf: medium
explicitReading AI-written documents in good faith requires senders to be able to defend each line; otherwise the document is a social violation ('slop').conf: high

Arguments

The era we are entering produces more, not less, human work — even as models become radically more capable. strength: high

Explicit premises:

Every agent currently needs a human who cares about it to remain useful (the 'reach test' and the maintenance burden of Open Claw)
Benchmarks measure performance on pre-articulated tasks but not the framing/prompting work that humans contribute
Each new model release commoditizes yesterday's competence and forces a new frontier of work
Empirically: Every doubled headcount in a year of intensive AI adoption

Implicit premises:

Human attention is the durable bottleneck and improving models do not relax it
The 'new frontier' work generated by each model release is paid work in identifiable roles — i.e. demand for human labour scales with model capability rather than being absorbed by leisure or by other automated agents
Every's experience (media-company / software-tools hybrid, ~30 curious self-selected early adopters with pre-release access) is representative of knowledge work writ large

Inferences:

If the bottleneck is human attention, the strategic constraint on enterprises is not compute or model cost but the supply of people who can ride the models — implying that the right organizational investment is in cultivating that population, not in agent count
But this inverts under Shipper's own argument once models genuinely reduce maintenance cost (a possibility he concedes in passing — 'as the models get better at being more independent... it'll be more likely we'll have more personal agents'). The 'more humans' claim is therefore conditioned on a current-tech window, not a structural law

Buy SaaS stocks; the SaaS-apocalypse story is wrong because agents increase SaaS demand and shift token costs to users.

Explicit premises:

Every's own SaaS spend is up year-over-year despite heavy agent usage
Agents make a high volume of calls into SaaS products (driving usage skyward)
When the user's agent (Codex) is the calling client, the user pays for tokens — protecting the SaaS vendor's margins
GitHub-style infrastructure load is already visible as evidence of agent-driven SaaS demand

Implicit premises:

Vendor pricing models will hold steady rather than be renegotiated to capture agent-driven value (e.g. per-call rate-limiting, agent-tier pricing)
Switching costs from agent-mediated SaaS to vibe-coded replacements remain high enough that displacement does not happen at scale
Codex/Co-work's role as token-payer remains user-friendly and is not absorbed back into vendor-billed consumption as enterprises seek cost control

Inferences:

If agents drive 10-100x request volumes against unchanged seat pricing, vendors will rationally migrate to usage-based pricing — at which point the 'tokens are user's problem' argument inverts and SaaS margin protection disappears
The strategic implication for SaaS founders is to bet on becoming the agent-readable substrate (CLI + good HTML), not on building proprietary in-app agents — Shipper says this explicitly but its inverse for incumbent vendors is unaddressed

PMs and full-stack designers are the role categories most positioned to thrive.

Explicit premises:

Coding models now let lightly-technical people build production software (Marcus / Spiral case)
Generative slop makes taste and judgment scarcer and more valuable
Designers can now ship interactions they previously had to hand off, reducing friction loss
Building is no longer the bottleneck; figuring-out-what-to-build is

Implicit premises:

PM and designer roles will retain their organizational legitimacy as autonomous shippers, rather than being absorbed into a broader 'product engineer' role that swallows both
Shipper's audience (PMs subscribing to Lenny's Podcast) is representative of the PM cohort that benefits — but the median enterprise PM, embedded in process-heavy organizations, may not

Inferences:

If 'figuring out what to build' is the new scarce skill, the rise of the PM as a winner has a darker corollary: PMs who are weak on that dimension (i.e. those who were valuable mainly as process coordinators, JIRA stewards, alignment facilitators) face the largest negative selection — Shipper does not flag this asymmetry

The right individual response is to 'ride the models': try each new release on your own work, play, find your moment of joy. strength: low

Explicit premises:

Frontier models are equally accessible to everyone with a credit card — geography and employer do not gate access
Models change rapidly, so what failed last release may work now ('turn the rock over')
Curiosity-driven exploration finds uses that lab researchers do not see ('the edge of AI is wherever AI meets a real human')
Fear- and FOMO-driven adoption produces worse intuition than play-driven adoption

Implicit premises:

Individuals will have the discretionary time, employer permission, and psychological surplus to play with frontier models — Shipper acknowledges some employers block this but treats it as soluble in personal off-hours
Play translates into job-relevant capability fast enough to outpace organizational competitive pressure
The cognitive style required to 'ride the models' (high openness, tolerance for failure, fast iteration) is broadly available rather than a personality-specific advantage

Inferences:

The advice is rationally individual but generates collective-action problems: if the workplace value of 'rode the model first' is winner-take-most, then aggregate welfare may decline (status competition among knowledge workers, longer hours) — consistent with Shipper's own paradox but in a register he does not surface

Narratives

Every doubled headcount from 15 to ~30 in the year since the last podcast appearance, while being one of the most AI-forward companies. Shipper frames this as evidence of, not contradiction to, the 'more automation, more humans, more work' thesis.
→ The single most important narrative — it operationalizes the unity statement and Shipper's headline paradox. Implicitly carries the load-bearing premise that hiring up is the right read of the situation rather than evidence of unrealized automation leverage
The senior-engineer benchmark: Shipper vibe-coded Proof (his markdown editor), it broke on launch, Codex couldn't fix it (he got bursitis trying — 'vibe coder elbow'), two human senior engineers independently rewrote it. Each new model is now scored against those rewrites. GPT-5.5 with an Opus 4.7 plan scored 62/100 versus humans at high-80s/low-90s.
→ Carries the 'benchmarks miss what matters' argument and, more importantly, the 'I will move the goalposts when it saturates' aside that betrays the implicit premise: human judgment is being defined as whatever AI cannot yet do, making the paradox unfalsifiable
Marcus, an Axios PM who took a year off to get AI-pilled, came to Every and now runs Spiral (writing app) shipping faster than almost anyone on the team, despite being only 'lightly technical'.
→ The single exemplar carrying the 'PMs win' prediction. Functions as proof-of-concept but is a maximally favourable case (year-long sabbatical, prior AI-pilling, strong product instincts) projected as a generalizable pattern
Shipper has been in inbox-zero for 10 straight days using Codex + Cora (Every's email agent) to render emails as a page he monologues through, with Codex doing research and drafting responses.
→ Concretizes the Codex-as-work-surface prediction; supports the AI-generated-writing-is-fine claim. Also slips in that he 'almost sent an investor email without looking' which is meant as anecdote but functions as evidence of the supervision burden Shipper's own thesis predicts
Recruiting anecdote: Shipper asked Codex to find someone with a General Assembly background who is AI-pilled to be Every's head of L&D; Codex found 'the perfect guy' who follows him on Twitter, leading to a DM and a dinner.
→ Supports sales/recruiting as use cases where AI augments rather than replaces; implicitly demonstrates the 'two agents' / agent-mediated SaaS pattern as already useful
Pete (Open Claw maintainer) receives thousands of pull requests a day and spins up '50,000 Codex instances' to sort and merge ~1,000. Shipper predicts this becomes common.
→ Illustrates the second-order overload problem when non-technical staff can also ship PRs — that capacity bottlenecks shift to review and integration
Notion-agent-driven quarterly planning at Every: each team talks to one Notion agent, which produces strategy docs, pushes back, relates plans to company strategy. Shipper then reviews the AI-generated docs to find inter-team dependencies.
→ Operationalizes the 'AI-generated writing is normal and welcome' proposition at the company-process level
Every's internal Claudie agent runs its whole consulting practice; AI engineer Nitesh spends most of his time talking to Claudie in Slack ('why did you do this dumb thing?')
→ Provides the concrete instance of the forward-deployed-engineer role and the super-agent organizational pattern
Customer support disappears: a user with a bug doesn't email — their agent files a bug report with exact repro steps, source-code hypotheses (Proof is open source), which becomes a GitHub issue another agent fixes.
→ Carries the two-agent-collaboration proposition and the SaaS-redesigned-for-agents prediction; implicitly assumes most users will use agent-mediated SaaS
The 'Middle Ages horizon' analogy: people imagine dragons or utopias beyond the horizon, but when you get there it's just another horizon with some good things and some bad things.
→ Rhetorical close that domesticates both AI doom and AI utopianism — supports the both-and posture of the unity statement

Questions addressed

✅ How will the way knowledge work happens change in the next 12 months?
Bifurcate into per-company super-agent (Slack) and Codex/Co-work as desktop work surface; SaaS persists but is increasingly accessed through an agent's embedded browser
✅ Will SaaS be killed by AI?
No; agents increase SaaS demand, users bring their own tokens, vendor margins are protected if vendors redesign for human+agent co-use
✅ How is it possible that AI-forward companies are still hiring up?
Every agent needs a human (forward-deployed engineers, supervisors); benchmarks understate human contribution; commoditizing yesterday's competence creates new frontier work — therefore aggregate human work grows
✅ Which roles thrive and which struggle?
PMs and full-stack designers thrive; sales / CEO / middle-manager roles change least so far; engineering/design/data science/PM craft change most; generalists beat specialists
✅ What should an individual do to stay employable?
Ride the models: try each new release on your own work, play, find your moment of joy, work in your off-hours if your employer blocks frontier access
✅ Are CLIs the future of agentic work?
No — the CLI era was brief; GUIs (Codex/Co-work) will dominate even for technical work
⭕ If every agent currently needs a human, what happens when models genuinely become independent?
Shipper concedes in passing that personal agents will return when maintenance burden drops, but does not engage with what this implies for his headline 'more humans' thesis if that transition arrives within 1-3 years
⭕ Does Every's experience generalize to the typical enterprise or non-tech knowledge-work organization?
Asserted by example; selection effects (curious self-selected early adopters, ~30 people, dual media+software business, pre-release model access) are not addressed
⭕ What about employees whose employers block frontier model access?
Briefly acknowledged ('a lot of companies handicap their employees'); only answer is 'do it on your off time' — no engagement with the equity/access asymmetry this creates between firms and within them
⭕ If the headcount-still-grows-with-AI claim is taken seriously, what happens to the standard CFO/board narrative that AI investment yields operating leverage?
Not addressed. Shipper conflates 'Every keeps hiring' with 'AI-native companies hire more' without disentangling growth-driven hiring (revenue scaling) from agent-supervision hiring (the structural claim)
⭕ What is the labor-market transition path for roles Shipper sees as 'most changed' — engineers, designers, data scientists, PMs whose previous craft is now commoditized?
Optimistic assertion that humans always move to the new frontier; no engagement with the speed, friction, or distributional consequences of the transition

Gaps

incompleteThe paradox is asserted but its mechanism is under-specified. Shipper offers 'every agent needs a human' (a maintenance claim), 'benchmarks miss prompting' (an attention claim), and 'models commoditize yesterday's competence' (a frontier claim). These are three different mechanisms doing different work, and Shipper conflates them. The maintenance claim is explicitly transient — he concedes personal agents return as models improve. The attention claim is partially circular — he will redesign his benchmark to zero out a saturated model. Only the frontier claim is structural, and it is the least developed. A serious version of the thesis needs to argue that the new frontier work is itself non-automatable and remunerated — neither is shown.
Read next: Acemoglu & Restrepo on task displacement and reinstatement; the 'augmentation vs. automation' literature (Brynjolfsson, Mitchell); recent econometric work on AI's effect on call-centre and software-engineering productivity (the GitHub Copilot RCT, Brynjolfsson-Li-Raymond on generative AI in customer support)
incompleteHeadcount as evidence is mis-read. Every grew 15 -> 30 in a year, but Every is a content-and-software company in active product expansion (Spiral, Cora, Claudie, Proof, a hosted Open Claw product, a consulting practice). The hiring may reflect revenue and product-scope growth — exactly what venture-backed startups do when their bet is working — not a structural inability to extract AI leverage. Shipper does not separate the two effects. For a senior engineering leader evaluating org impact, this matters: the AI-paradox-driven hiring claim and the new-business-line-driven hiring claim have opposite implications for whether enterprises in steady-state functions can shrink.
Read next: Contrast with the O'Reilly 'Agentic P&L: Beyond the Empire of Headcount' extraction in this same library (extract-agentic-pnl-beyond-empire-headcount.json) — Shyamsundar & Jain make a structural case that headcount is a 'liability' and that compliance functions shrink 4-5x. Shipper and they cannot both be right about the same situation; the disagreement is precisely about whether residual hiring after AI is (a) the supervision tax Shipper sees, or (b) leftover political capital the Agentic P&L authors expect to dissolve. The likely synthesis: Shipper is describing an early-adopter product company at growth stage, while Agentic P&L describes a steady-state enterprise function — both are true but neither is generalizable to the other case. Read alongside Aaron Levie's enterprise-cost-curve essays and recent McKinsey / BCG empirical surveys of AI deployment outcomes 2025-26
closed byAI Companies Are Hiring More — Provides the cross-firm evidence the gap demands: matched-firm Ramp/Revelio data showing high-AI-adoption firms grew headcount ~10% (entry-level ~12%) versus flat low-adopters, generalizing the hiring-up claim beyond the single Every case (causality caveats remain).
incompleteThe 'humans-as-orchestrators' vision is unfalsifiable in this conversation. Shipper says explicitly that when models 'saturate' his senior-engineer benchmark, he will redesign it to zero out the current model. If the benchmark moves to wherever AI cannot yet go, the claim 'humans are still needed' is true by construction and tells the listener nothing about the actual trajectory of automatable knowledge work. A senior engineering leader needs a falsifiable formulation — what specifically would need to be true for Shipper to concede that the 'more humans' claim has expired? — and the conversation does not offer one.
Read next: Popper on falsifiability; Sutton's 'bitter lesson' as the counter-frame; the eval-design literature (HELM, MMLU saturations, the ARC-AGI debate); recent critiques of human-baseline benchmarks (Anthropic's METR work on long-horizon agentic tasks)
incompleteSelection effects in the Every case are uninterrogated. Every is ~30 self-selected curious early adopters with pre-release model access, dual media-and-software business model, founder writer-CEO, no legacy systems, no regulated environment, no enterprise sales process longer than the model release cycle. Every assertion in the conversation is sourced from this one organization and projected onto knowledge work in general. For evaluating organizational impact in a Norwegian enterprise context (regulated industries, slower procurement, less English-language model access for non-English work, strong worker-council constraints), the generalization gap is large and unflagged.
Read next: Erik Brynjolfsson's recent work on cross-firm productivity dispersion in AI adoption; Wharton-led surveys on enterprise AI deployment; case studies from regulated EU firms (Klarna's published numbers and their later retraction); Norwegian-context: BI Handelshøyskolen / Sintef studies on AI adoption in Nordic enterprise (the 'forskning og digitalisering' extraction in this library)
incompleteToken-economics assumption (SaaS users bring their own tokens via Codex) is treated as a stable equilibrium when it is more plausibly a transitional artifact. Once enterprise procurement notices that token spend is shifted to user-attached agent harnesses, three forces will push back: (a) per-call agent-tier pricing from SaaS vendors capturing the new volume, (b) enterprise consolidation of agent infrastructure (corporate-controlled Codex/Claude Code deployments), and (c) usage-based SaaS pricing emerging precisely to capture agent-driven volumes. Each of these inverts Shipper's 'buy SaaS stocks' conclusion. He does not engage any of them.
Read next: Tomasz Tunguz's recent essays on AI pricing; the cloud-pricing history (AWS's evolution from flat to consumption); Bessemer State of the Cloud reports on agent-driven pricing changes; OpenRouter / LLM-aggregator economics
incompleteShipper's 'CEO must be personally AI-pilled' claim sits uncomfortably with his 'super-agent per company maintained by a forward-deployed engineer' claim. If the FDE is the load-bearing role and the per-company agent is the substrate, the CEO's personal model fluency is arguably the wrong variable to optimize — the firm needs FDE talent supply and an org design that protects FDE attention from the C-suite. The conversation never reconciles the personal-fluency prescription with the structural FDE prescription.
Read next: Karl Weick on 'sense-making' in organizations; recent CIO/CTO surveys on AI organizational placement (Gartner, MIT Sloan); the literature on 'two-pizza team' analogues for AI deployment
incompleteNo engagement with worker-attention costs of the prescription. 'Ride the models' for every new release, 'try Codex in your off-hours if your employer blocks it', 'every agent needs a human gardener', 'inbox-zero for 10 straight days via Codex monologue' — collectively this is a maximally attentive lifestyle marketed as liberation. Shipper notes he 'works way more' and frames it as paradox-evidence, not as cost. For a senior engineering leader thinking about sustainable organizational design, the elision of attentional and psychological cost is significant.
Read next: The 'AI psychological costs' extraction in this library (extract-psychological-costs-of-ai-*); Cal Newport on attention residue and deep work; recent labour-economics work on AI-driven work intensification (LSE, Stanford); union/works-council positions on AI workload
illogicalThere is internal tension between (a) 'CLIs are over, GUIs win' and (b) 'agents using SaaS via CLI is how SaaS survives'. If the user's primary surface is Codex/Co-work GUI and the agent talks to SaaS over CLI, then CLIs are precisely what is preserved — they are the agent-readable API layer, not the user surface. Shipper conflates the user-surface CLI claim with the agent-readable-API CLI claim. The two are different and have opposite trajectories: the first dies, the second flourishes. The conversation does not separate them.

The AI paradox: More automation, more humans, more work | Dan Shipper

Structure

Key terms

Propositions

Arguments

Narratives

Questions addressed

Gaps

Related documents