Lars Faye's Agentic Coding Is a Trap — published Sunday, May 3, picked up on Hacker News at 398 points and 316 comments — is the best single compendium of the cognitive-debt evidence base anyone has put together in 2026. It catalogues the studies. It names the trade-offs. It lands on a personal-discipline conclusion. The receipts are now collected; the careful reader will have spent the weekend nodding through them.

Buried in Faye's second paragraph, almost in passing, is the line that does the actual analytical work. Faye describes the agentic workflow as a process in which "someone defines the project's requirements ... generates a plan, and then pulls the slot machine lever over and over, iterating and reiterating with often multiple agent instances until it's done." The link goes to a March post by Quentin Rousseau, CTO and co-founder of Rootly, titled One More Prompt: The Dopamine Trap of Agentic Coding. The metaphor isn't Faye's. Rousseau got there first, in clinical language: the workflow runs on "variable ratio reinforcement — the same psychological mechanism that makes slot machines the most addictive form of gambling".

That is the framing the rest of Faye's piece is downstream of, and it is the framing this article is about.

What the receipts add up to

Faye's catalogue, briefly. Anthropic's own research note on internal use names what it calls the "paradox of supervision": effective use of Claude requires the very skills that sustained Claude use atrophies. MIT Media Lab's Your Brain on ChatGPT measured the cognitive impact and labelled it cognitive debt. A Microsoft study covered by 404 Media reached parallel findings for knowledge workers more broadly. A separate Anthropic study on coding skills reported a 47% drop-off in debugging skills among engineers leaning heavily on AI-assisted workflows. Sandor Nyako, the LinkedIn engineering director who oversees fifty engineers, has reportedly asked his team not to use these tools for "tasks that require critical thinking or problem-solving."

These are well-credentialed studies, performed mostly by parties with no incentive to overstate the effect. Each one names some symptom: cognitive debt, debugging atrophy, skill-formation interruption, supervisory paradox. The piece this article is responding to has done the hard work of collecting them.

What the catalogue underspecifies is the upstream question. Why does this particular workflow produce these particular symptoms? The answer is in the link Faye's second paragraph throws away.

What Rousseau actually said

Rousseau's March post is unusually direct. The author, writing as a working CTO of an early-stage company, names the workflow's reward schedule and its physiological consequences in the same paragraph. The agentic-coding loop, in Rousseau's account, is structured around intermittent reinforcement. Sometimes the diff is what you wanted, sometimes not, sometimes spectacularly close, sometimes laughably wrong. The "intermittent reinforcement of those dopamine and adrenaline hits creates the core addictive pull," in Rousseau's phrasing. The behaviour the schedule produces, in Rousseau's reporting from the Y Combinator founder community he is part of: developers "routinely coding until 2-4 AM despite no deadline pressure", the author himself reaching for orexin-receptor-blocker prescriptions to push back against the wakefulness effect, and a public comparison from Garry Tan describing the dopamine return as comparable to manually finding the answer. Rousseau also reports that approximately 25% of the most recent Y Combinator batch has codebases described as "almost entirely AI-generated".

This is the framing Faye is referring to, and it is not metaphorical decoration. The engineering-cohort observation is that a particular workflow produces a particular reward schedule, and that reward schedule produces a particular pattern of behaviour, including pharmaceutical countermeasures. The behaviour pattern is not coincidence. It is the engineered output of the loop.

What the workflow is shaped for

If the workflow's reward schedule is variable-ratio reinforcement, the question is whose problem that solves. The engineer's problem is that the work needs to get done. The vendor's problem is that the engineer needs to keep paying for tokens. The two problems do not point in the same direction; one of them gets solved more thoroughly than the other.

Faye's piece links to reporting on a related dynamic: AI adoption inside organisations is being measured in tokens spent, and that measurement is being used as a proxy for productivity. Token count is the easiest number for an engineering-management dashboard to render; it is also the revenue line item for the vendor. The metric and the price of revenue are the same number, which is unusual, and worth thinking about. The Uber data published earlier this month, with per-engineer monthly token bills running to $500–$2,000, the engineering organisation ramping from 32% to 84% adoption in four months, and the entire 2026 AI budget consumed in the first quarter, is the corporate-finance-line-item version of the YC founders Rousseau describes coding to 2 AM. The lever is the same lever; only the cadence and the venue differ. Each engineer pulling it at industrial frequency is one row in a budget the CFO did not anticipate.

The alignment is not pedagogical. It is industrial. It is the same alignment that produced the previous decade's attention economy, with the engineer in the seat the social-media user used to occupy.

We have done this before

The historical analog is not assembly-to-FORTRAN, the comparison Faye explicitly rejects in his piece, and rejects correctly. "a higher level of ambiguity is not a higher level of abstraction," in Faye's phrasing, and the FORTRAN frame flatters the new tools by aligning them with a pedigree of advances they do not earn. The honest analog is closer to home, in the same fifteen-year window many readers of this piece have lived through.

The compression of the recognition window is the part most worth noticing. The attention-economy harms took a decade to accumulate enough peer-reviewed evidence to argue about; the token-economy harms have a paradox-of-supervision admission from the largest vendor inside three years. The cohort doing the measurement also happens to be the cohort being measured, which speeds the reporting.

What the lever pulls cost, one engineer at a time

The HN thread on Faye's piece is unusually heavy on testimony from inside the senior bracket. The senior-engineer-cannot-answer-questions scene that the previously-published companion piece What We Lose When Coding Becomes Reviewing centred is one such datapoint; what concerns this piece is the moment immediately downstream, when the same engineer reaches for the same workflow again the next morning. One commenter with thirty-five years of experience offered a more cheerful counter, that agentic tools had let them learn more in the last few years than in the prior thirty-five, only to draw an immediate reply that this is a curve available only to engineers who already had thirty-five years of friction in the bank to draw on. Both readings can be right. The point one of them was making, deeper in the same comment thread, is the one that keeps catching: "I think a great deal of what made computing an amazing industry to work in is going to or has already died." Whether the speaker is right depends on how the next five years go. The reading is not a complaint; it is a description offered without satisfaction by someone who watched the previous version.

What the lever pulls cost the individual engineer, in the cases the studies are now measuring, is the cognitive practice that produced the engineer in the first place. The slot-machine analogy is exact in the wrong way: a casino visitor leaves with thinner pockets and the same brain. The agentic-coding loop costs the brain.

Coda

The slot-machine framing is not a complaint. It is a description offered, not for the first time, by people who have noticed that the workflow's reward shape and the vendor's revenue shape are the same shape, and that the alignment has consequences. We have done this once before, with a different captive population and a different metering surface, and the consequences took a decade to be argued about with a straight face. The compressed timeline this time is a small mercy. The receipts arrived faster. The remaining question is whether the recognition is going to do any structural work, or whether the field, having decided that demote AI's role is a sufficient answer at the individual level, will accept that as the answer at the institutional level too. The cost was not a bug. The cost was the design. Every previous case of this pattern was eventually answered by someone with the standing to write a rule about it. The slot-machine industry, eventually, accepted some.