The capitalist must concentrate in his own hands or his managers’ the knowledge that the workers formerly possessed.

Harry Braverman, Labor and Monopoly Capital (1974)

The Stage

“There’s no force on this planet that can make that LLM solve this problem for you.”

The line is Andrej Karpathy’s. The stage is Sequoia Capital’s AI Ascent conference. The date is April 29, 2026. He is on stage with partner Stephanie Zhan, in front of the largest concentration of AI capital in the room, and what he is doing for the next forty-five minutes is burying the phrase he coined exactly one year ago: “vibe coding.”

The talk is titled From Vibe Coding to Agentic Engineering. The vocabulary shift is the entire content of the talk. Vibe coding was the 2025 marketing moment, when capital needed retail engineers to believe that any twenty-year-old with a Cursor subscription was now a senior engineer. “Agentic engineering” is the 2026 industrial moment, when capital needs the same engineers to believe that producing software now requires a permanent class of credentialed human handlers paid by frontier labs and embedded inside customer accounts. The vibes have been industrialized.

The codebases of every firm that ran the first wave are now privately enclosed inside four hyperscaler clouds (a process this blog named in entry #9, The $725 Billion Enclosure). Enclosure, in the historical sense, is the conversion of common land into private property and the eviction of the people who used to work it. Software is now enclosed in this sense. The four firms that own the inference do not own the labor that produced the training data, but they own the choke point through which that labor must now pass to reach a market. The choke point is the asset.

Karpathy does not put it that way. Karpathy is too smart and too honest for the boosters’ standard liturgy. What he says, on stage, is a series of admissions. The admissions are perfect. The admissions are the new marketing.

The Heart Attack

Karpathy says, on stage, that when he looks at the code his agents produce, he gets a heart attack.

Quote, lightly cleaned. “When you actually look at the code, sometimes I get a little bit of a heart attack because it’s not super amazing code necessarily all the time. It’s very bloaty. There’s a lot of copy-paste. There are awkward abstractions that are brittle. It works, but it’s just really gross.”

This is the frontier. This is the output of a multi-hundred-billion-dollar capital cycle, narrated by one of its principal architects, on the stage of the venture fund that financed most of it. It works. It is also gross. It gives Karpathy palpitations. Brittle, bloaty, copy-paste, awkward. The model has been trained at a cost that would have funded a generation of state housing, and what it produces is “really gross” code that requires a human cardiac event to ship safely.

If this is the frontier, every layoff memo written by a CEO citing AI productivity gains in the last twelve months is, in some non-trivial sense, fraud. Not literary fraud. Securities fraud of the polite kind, where the misstatement lives in the productivity inference rather than the cash-flow disclosure. The Coinbase memo, the Cloudflare memo, the IBM memo, the Salesforce memo. Each of them sold investors a story that the model is replacing labor. The model is producing code that puts its own designer at risk of a coronary event. The labor was not replaced. The labor was rebranded as supervisory cognition and offered back to the worker at sixty percent of the previous wage.

The fiscal logic of the rebrand is clean. If the AI is doing the work, the supervisor of the AI is performing a less skilled task than the worker the AI replaced, and the supervisor’s wage should reflect the reduced skill input. This is the argument the CFO makes to the board. The argument is not true. The supervisor is performing a more skilled task than the worker the AI replaced, because the supervisor is now responsible for catching the catastrophic edge case the model will, eventually, generate. The wage cut is justified by a fiction. The fiction is sustained by the productivity discourse the same boosters who admit to coronary events on the Sequoia stage refresh quarterly.

Stopped Checking

Earlier in the same talk, Karpathy describes a different progression. He used to check the agent’s code. He used to correct it constantly. Now he has stopped. The models, he says, have improved to the point where he lets the output ship. He frames this as a productivity tip.

This is twenty minutes of stage time apart from the heart attack confession. Both statements are true. They are simultaneously true. Karpathy has stopped checking the code AND Karpathy has a coronary event every time he checks the code. The reconciliation is obvious once you stop performing the boosters’ liturgy. He has stopped checking the code most of the time, and the bad code is shipping. He checks it occasionally, and the bad code is, in fact, bad enough to put him in arrhythmia. The cognitive labor process Karpathy is performing is: skim, defer, panic, fix, ship. The narrative says “autonomous agent.” The practice is anxiety-ridden supervision punctuated by selective despair.

This is the labor process now. It is not described in any quarterly earnings report. It is the gap between the productivity claim and the production reality. The Forward Deployed Engineer fills it for enterprise customers (see Forward Deployed, #27 in this series). Karpathy fills it for himself. The Coinbase remainder workforce, the Cloudflare survivors, the IBM operations team, they all fill it for their respective employers, and they fill it at a wage that has been reset downward on the explicit theory that the AI is doing the work.

Translate. The AI is not doing the work. The AI is doing the typing. The work has been relocated, partially socialized to the user, and stripped of its rent.

Three words load that sentence. Relocated: the cognitive labor of producing working software has not been eliminated. It has moved. From writing code to writing specs, supervising output, catching catastrophic edge cases, refactoring the bloaty copy-paste the agent emits. Karpathy describes this workflow himself.

Socialized to the user: offloaded onto the broader user base rather than performed by paid specialists inside a firm. IKEA sells you a box of parts. You provide the assembly labor for free. The labor cost is socialized to the customer and stripped out of the price. The AI coding case is closer to IKEA than to anything else. The frontier lab sells you an agent. You provide the spec-writing, the supervision, the verification, and the fix-up labor. That labor used to be performed by paid engineers inside firms and billed at engineering wages. It is now performed by you, the subscription holder, often unpaid, often paid at a wage that has been reset downward on the theory that the AI is doing the work.

Rent: in the classical economic sense. Income that flows to the owner of a scarce asset by virtue of ownership, not production. The landlord takes rent on the field he did not till. The platform takes rent on the marketplace it did not stock. The model owner takes rent on the cognitive labor the engineer still performs but no longer captures. The wage premium a senior engineer used to earn was the engineer’s slice of the rent on the scarcity of working-software production. The first move of the current cycle is to declare that scarcity ended. The second move is to reroute the rent to whoever owns the model and the services contract. The engineer still does the work. The engineer no longer collects the rent on doing the work.

The Strawberry

Karpathy is, to his credit, the most articulate ghost-whisperer in the booster wing. He has stopped calling the models animals. He calls them ghosts now. Jagged, statistical, summoned entities. The metaphor is correct. The metaphor is also the new commercial scaffolding.

He gives the audience the example everyone in the room has heard. The same model that refactors a 100,000-line codebase cannot reliably count the letter R in the word “strawberry.” It cannot reliably decide whether you should drive or walk to a car wash. The skills are jagged. The technology that is supposedly replacing doctors, lawyers, and accountants cannot count letters in a ten-letter common noun.

The jaggedness is the actual technical fact. The marketing requires the smoothed version. The “AI Engineer” job posting on LinkedIn promises domain-replacing capability. The Karpathy talk admits domain-bounded capability that fails outside the reinforcement-learning distribution. He says this directly, on stage. “If the task you’re trying to do is not well represented either in the base data or in the RL data, there’s no force on this planet that can make that LLM solve this problem for you.” No force on this planet. Said at Sequoia.

The Menu Gen anecdote lands the practical point. Karpathy is building a side project. The agent makes an “assumption that just didn’t make any sense,” reusing emails in a way that would have been “really catastrophic” if Karpathy had not caught it. He tells this as a charming personal anecdote. There is nothing charming about it. This is the same class of model being deployed into healthcare triage workflows, insurance claims adjudication, and judicial sentencing recommendation systems. The charming personal anecdote is the operational risk profile.

The booster move is to call this a “skill issue.” The user did not specify well enough. The user did not prompt correctly. The user did not write the spec with sufficient detail. The cognitive labor of catching the catastrophic edge case has been silently transferred to the user, and then renamed as the user’s deficiency when the catch fails. This is what wage discipline sounds like in a TED-adjacent register. Wage discipline, briefly: the mechanism by which an employer convinces a workforce that the wage on offer is the wage they deserve, usually by reframing structural failure as individual deficiency. The “skill issue” diagnosis is wage discipline in three syllables.

The Spec

Karpathy describes his actual workflow. He writes “an extremely detailed document explaining everything that you want the code to do. Every single edge case, you write it in markdown in English, and you try to be as precise as possible.” Then the AI generates the code from the document.

This is coding. Writing an extremely detailed, precise, complete specification with every edge case named is the act of programming. The medium has shifted from one formal language (Python, TypeScript, Rust) to another formal language (English-disguised-as-prose, structured by markdown). The cognitive labor has not decreased. It has been displaced sideways and is now performed in a register that is harder to credential, harder to bill, and harder to unionize around. The engineer is still solving the problem. The engineer is now also performing the solution as a markdown document while pretending the document is a request.

Harry Braverman called this the separation of conception from execution. The early industrial capitalist’s first job was to take the craft knowledge that lived in the worker’s hands and move it into the manager’s clipboard. Conception was concentrated upstairs. Execution was deskilled downstairs. The wage was reset on the new theory that the worker was now “just” executing. Spec-writing is the conception. The agent is the execution. The engineer who used to do both is now charged with the harder half, paid for the easier half, and told that the AI is doing the work.

The Recantation

The vibe coding pivot is the part of the talk that matters most for capital allocation.

A year ago, vibe coding was the marketing apparatus. Anyone could ship software now. The barrier to entry had collapsed. Capital flowed into AI coding tools on the implicit promise that the labor input could be replaced with subscription revenue. The narrative produced its layoff cycle. The layoff cycle produced its enclosure of the codebases and its concentration of compute spend in four hyperscalers. The enclosure is permanent.

A year later, Karpathy is on the same stage telling Sequoia’s audience that vibe coding was the on-ramp. Agentic engineering is the destination. Agentic engineering requires “taste and judgment,” a “new kind” of skill, specifications, supervision, verifiability tooling, and the integration work that frontier labs are now standing up enterprise services firms to provide. The Anthropic / Blackstone / Goldman Sachs / Hellman & Friedman / Apollo / General Atlantic / Leonard Green / GIC / Sequoia consortium announced May 4 is the institutional form of this pivot. The Forward Deployed Engineer is its job title.

Translate. The first wave sold the disappearance of labor to justify the capex. The second wave sells the indispensability of a new credentialed labor caste to monetize the gap the first wave left behind. The vocabulary shift from “vibe” to “agentic” is the rent shift from “AI tool subscription” to “AI services contract.” The same capital that financed the first wave is now collecting on the second. Karpathy’s talk is the technical scaffolding for the move. The audience in the room is not the audience for the talk. The audience for the talk is the LPs of the funds in the room, who need a story about why the second cheque is rationally underwritten by the same thesis that justified the first. The story is that the technology is real but the human is necessary. Both halves of that story are deployed for billing purposes.

He does this without lying. He does this by being honest about jagged skills, verifiability gaps, base distributions, and the heart attack he gets when he reads the output. The honest booster does not break the apparatus. The honest booster is the apparatus. Gramsci called this transformismo: the political technique by which a regime absorbs its most credible critics into its own management layer, neutralizing the critique by promoting the critic. The most articulate internal skeptic does not destabilize the structure. The most articulate internal skeptic becomes the philosophical license for the next round of accumulation. Karpathy’s openness is the asset. The asset is being deployed.

Refuse

The pamphlet form demands a program. Four moves.

Name spec-writing as engineering labor. The wage reset on the current cycle depends on the fiction that writing the markdown document is a request while writing the Python is work. Refuse the distinction. The spec is the program. The spec-writer is the programmer. Bill the spec-writing hours. Track them. Put them in the performance review. Put them in the job listing. If the firm wants to call you an agentic engineer, the firm pays for the agentic engineering, which is the part where you write the specification and supervise the ghost. The pay scale follows the labor process. The labor process is upstream of the prompt.

Refuse the vocabulary churn. Vibe coding lasted one year before recantation. Agentic engineering will last roughly the same. The 2027 term will arrive on schedule. Each rebrand is a wage discipline event dressed as a taxonomy update. Each one is calibrated to make the previous cohort’s skills look obsolete and the previous cohort’s wage look generous. Refuse to learn the new word until the labor process behind the old word has been honestly accounted for. The taxonomy is downstream of the rent. The rent is the structure.

Solidarity across the supposedly-replaced and the supposedly-upskilled. The laid-off Coinbase support agent and the senior engineer doing spec-writing at the surviving Cloudflare are not in different stories. They are in the same labor process, processed by the same apparatus, at two different stations on the same line. The wage cut on the first funds the rent on the second. The booster narrative is built to keep them from recognizing each other, because the recognition is the political risk. The senior engineer who thinks the support agent’s job loss is unrelated to her own wage compression is the senior engineer the apparatus is counting on.

Demand public compute and model auditability. The black-box ghost is part of the rent extraction structure. If the model owner cannot show you what is inside the model, what data it was trained on, and how the reinforcement learning was conducted, the rent the model owner collects on its output is unaccountable in the strict sense: it cannot be audited. Public infrastructure for inference and training, even at a small fraction of the hyperscaler footprint, breaks the enclosure. This is not a maximalist demand. This is the floor.

None of these moves block the technology. The technology is real. The capability is partial. The argument is not against the model. The argument is against the wage structure being built on top of the model under the cover story that the model has done the work.

Skill Issue

Andrej does not have answers. The interviewer asks what is worth learning if AI gets good. Karpathy gives an empty answer. This is not because Karpathy cannot answer. It is because these are hard questions that nobody has figured out, “even Andre.”

This is the entire situation, named honestly. The most technically literate, financially incentivized, institutionally embedded, publicly articulate booster of the current AI moment cannot tell you what to learn, what to build, or how to insulate yourself from the labor process restructuring his industry is executing on the rest of the economy. He has a heart attack when he reads the code. He has stopped checking the code. He calls the models ghosts. He writes specs in English and pretends the specs are requests. He gives an empty answer to the question of what comes next. He cashes the consulting cheque on the way out.

His incentive is to keep the rent flowing. His honesty is the form the rent now takes. The pattern is not new. The English wool merchants who enclosed the commons and then wrote moral tracts about the laziness of the displaced peasants. The railway financiers who immiserated a generation of track workers and then sponsored libraries to commemorate the achievement. The postwar mortgage securitizers who designed redlining and then funded urban renewal think tanks to study the resulting poverty. The platform consolidators who hollowed out local journalism and then endowed chairs in media studies at the universities that trained the replacement gig workers. Each generation produces its honest internal narrator. The narrator’s honesty is part of the deal. The narrator gets to keep his stake. The workers being processed by the apparatus the narrator describes do not get to keep theirs.

The cohort the apparatus is currently processing is the one being told the failure is a “skill issue.” The senior engineer who cannot one-shot the Twitter clone with cookie expiration handled. The junior engineer who has not yet learned to write specs. The laid-off Coinbase support agent whose entire job has been replaced by a model that cannot count letters in “strawberry.” The mid-career project manager being retrained for “AI fluency” while the firm uses the retraining period as cover for the wage reset. Each of them is being told, by a productivity discourse refined and re-released at every venture conference, that the failure is theirs.

The failure is not theirs. The failure is the gap between what was sold and what was built, monetized as a consultancy bill on the way down and as a wage cut on the way through. The model is real. The capability is partial. The press release is fluent. The Forward Deployed Engineer is hired. The workers are gone. The investors are paid.

This blog has, over a few entries, named the principals, the mechanism, and the political economy. This entry adds the cognitive labor process. The next twelve months will not be decided by model capability. The next twelve months will be decided by whether the people performing the spec-writing, the supervision, the verification, the integration, and the fix-up labor of this technology recognize that they are doing the work, and whether they can name it before the rebrand from “vibe” to “agentic” to whatever-the-2027-vocabulary-turns-out-to-be locks the wage structure in.

Skill issue is the diagnosis the apparatus offers. The diagnosis is wrong. The skill is fine. The issue is the rent.