Naming the Loop: The Hidden Cost of Agentic Coding (Before We Learn How to Break It)

The previous essay in this small series was about builders being displaced — about a world that, in the early years of agentic AI, looked as though it might no longer have a place for people whose pleasure is in making things. It was a worry from the outside in. What happens to us if we have nowhere to build.

This essay is the dark mirror. What happens to us when we have everywhere to build, all the time, with zero friction. Same population, different threat. The first asked where the energy goes when the doors close. The second asks what the energy does to us when the doors are open at every hour, in every direction, with a slot machine behind each one.

The shape of the second worry is best caught in the words of a senior engineering manager, interviewed in a recent Boston Consulting Group study¹:

"I had one tool helping me weigh technical decisions, another spitting out drafts and summaries, and I kept bouncing between them, double-checking every little thing. But instead of moving faster, my brain just started to feel cluttered. Not physically tired, just… crowded. It was like I had a dozen browser tabs open in my head, all fighting for attention. I caught myself rereading the same stuff, second-guessing way more than usual, and getting weirdly impatient. My thinking wasn't broken, just noisy — like mental static. What finally snapped me out of it was realizing I was working harder to manage the tools than to actually solve the problem."

The Scene

There is a scene in Vieni avanti cretino, a 1982 comedy with Lino Banfi², in which the protagonist begins a new office job. The manager explains the task. At the first beep, press the red light. At the second beep, press the green light. At the third beep, turn everything off. Simple, the manager says. Very simple. And: your satisfaction is our greatest reward.

A minute later the manager comes back. Between the red and the green, would you mind also activating the telex? Press Q to start, S to stop, eight holes to punch. Simple. Very simple.

He comes back again. While you're doing that, could you also answer the phone? But before you answer, pull this lever to activate Circuit Z, then say "Dr. Thomas is not available." Very simple.

And again. When you hear the acoustic whistle, blow into it, then pull this other lever to activate Circuit Y. Very simple.

By the end of the scene the protagonist is babbling — red green penalty kick out of office no satisfaction corrupt referee — confusing the manager's instructions with a football match playing somewhere in his head, his brain having given up on the difference between his various inputs. The final image: he asks for a broom, because his backside is the only part of him not yet performing a task.

The comedy is in the buttons. The interesting part is in the manager.

What is striking, watching the scene again, is what the manager never does. He never raises his voice. He never imposes a deadline. He never says you must. Every new task arrives with the same polite preface: as long as you don't mind, as long as it doesn't bother you. The protagonist always says yes. The collapse is not a speed-up. It is the cumulative weight of a hundred consents to small additions, each one trivial in isolation.

For decades the scene has been read as a joke about industrial labour, in the lineage of Chaplin in Modern Times. That reading is fine, but it misses what is most disturbing about it. The protagonist's manager is not a factory. The factory floor in Modern Times speeds the conveyor against the worker's will. The manager in our scene does the opposite: he asks. He invites. He keeps the protagonist's consent at every step. The horror is that consent, repeated often enough at low enough stakes, produces exactly the same outcome as coercion.

This is, more or less precisely, the situation of an engineer using AI agents in 2026 — with one twist that matters.

It is tempting to cast the agent in the role of the manager. To say: the agent keeps asking, politely, for one more task. The framing is almost natural. Would you also like me to refactor this module? Should I add a test? I noticed a related bug — want me to fix it while I'm here? It sounds like asking. It is not.

The agent does not, strictly speaking, want anything. It produces output. We read the output as invitation. We open the next prompt. The asking happens inside us, and the answering happens inside us, and the agent is just the cheapest way to keep the cycle running.

This matters because it relocates the problem. The polite manager is not in the room — we are the polite manager. AI compresses our tasks and leaves small pockets of silence in the day. The thirty seconds while the agent runs. The five minutes between a passing test and the next ticket. The hour that used to go to thinking before keyboard-touching. We cannot stand these pockets. We fill them. One more prompt. One more refactor. One more worktree open. Every yes is one we said to ourselves.

No PM forces anyone to take more tickets. No leadership memo demands fifteen story points where ten used to live. The backlog discovers your new capacity faster than you do, but only because you offered it. Closing a ticket feels good. Taking on one more — the small, doable one, the one you can probably knock out before lunch — feels better. The dashboard notices. So does your own quiet comparison to last sprint. You take on more, and nobody asked you to.

By eleven at night there are six worktrees open and a creeping suspicion that the broom is the only thing left.

AI Brain Fry

The pattern has started to surface, awkwardly, in the writing of senior engineers. Francesco Bonacci, who runs an open-source agents company in San Francisco, named it earlier this year: vibe coding paralysis³. He described ending his days exhausted not by the work itself but by the managing of the work — six worktrees open, four half-written features, two quick fixes that spawned rabbit holes, and a creeping sense of losing the plot entirely. Steve Yegge, more dramatically, launched a tool called Gas Town on New Year's Day, an open-source platform for orchestrating swarms of Claude Code agents simultaneously⁴. One early user described it as moving too fast for him to comprehend, with a palpable sense of stress.

These are loud cases. There are quieter ones. In a recent Tom's Hardware Italia piece⁵, an Italian researcher described workers firing off AI requests during the dead moments of their day — before lunch, while making coffee, on the commute. Not because anyone asked them to. Because the tool was there and the moments felt wasteful.

The Boston Consulting Group surveyed 1,488 full-time US workers in January and found that fourteen percent of AI users at large companies report a syndrome they have named AI brain fry: mental fatigue from excessive use or oversight of AI tools beyond one's cognitive capacity¹. In marketing the figure reaches twenty-six percent. In engineering, eighteen. The descriptions in the study are remarkably consistent across roles: fog, buzzing, mental static, a dozen browser tabs open in one's head.

The pattern is structural, not individual. It is not Yegge being Yegge. It is the broad middle of knowledge work running into something the productivity discourse never warned us about.

Whiplash, Not Modern Times

The scene gives us the structure of the trap — small consensual additions, accumulating until the system collapses. What it does not give us is the psychology of why the protagonist keeps saying yes. For that, the film to reach for is Whiplash⁶.

Andrew Neiman bleeds on his drum kit because he believes he is purchasing greatness with the bleeding. The transaction feels honourable: pain now, transcendence later. Damien Chazelle's film never resolves whether he is right. The final scene is genuinely transcendent and genuinely pathological, simultaneously, and the refusal to choose between the two readings is the horror of the picture.

This is the correct frame for what is happening to engineers with agentic AI. The trap is not that the activity is fake. The trap is that it is real. You do ship code. You do close tickets. You do learn things. The realness is what makes the loop unbreakable. A pure waste of time would be easy to quit. The reason the protagonist keeps consenting to every new task is that each task is, in itself, a small genuine win. The reason an engineer keeps approving the agent's polite suggestions is the same. The seesaw tips toward pleasure on every yes.

The Dopamine Mechanism

To understand what is actually happening, it helps to read Anna Lembke, the Stanford psychiatrist who runs the addiction medicine clinic and wrote Dopamine Nation in 2021⁷. Lembke does not write about AI. She writes about smartphones, social media, gambling, food, pornography. Her core model, drawn from a 1974 paper by Solomon and Corbit⁸, is the pleasure-pain balance.

The model is mechanical. Pleasure and pain sit on opposite sides of a literal seesaw in the brain. When a stimulus tips the seesaw toward pleasure, the brain does not simply enjoy the tilt. It actively pushes back, restoring equilibrium. The pushback is felt as a small dysphoria — a low, flat moment after the high. Crucially, the pushback rises more slowly than the pleasure spike, but it lasts longer.

This is unremarkable for a single event. A drink, a slice of cake, a notification — pleasure peaks, dysphoria follows, baseline returns. The mechanism only becomes pathological when the stimulus is repeated often enough that the seesaw never returns to neutral between events. Under chronic stimulation, the pain side strengthens; the pleasure side shortens; eventually the baseline itself shifts downward. Lembke's clinical patients describe this as the moment when they no longer use the substance to feel good — they use it to escape the pain of not using it.

The variable that matters, Lembke is careful to point out, is not intensity. It is frequency. A Saturday night with friends and wine does not tip the system. A hundred prompt-response cycles a day, every day, for six months, does.

It is not difficult to see why this might apply to agentic coding. Every prompt is a small pleasure spike — a task closed, an artifact produced, a problem briefly handled. Every pause between prompts is a small dysphoric pushback. By the evening, the only thing that relieves the accumulated dysphoria is another prompt. By the next morning, the seesaw has not fully returned to neutral. The day starts a click lower than the one before. This is what the engineering manager in the BCG study is describing when he talks about static and crowdedness and working harder to manage the tools than to solve the problem. He is not lazy. He is not undisciplined. His pleasure-pain balance has been recalibrated by hundreds of low-grade hits a day, and the recalibration is now interfering with his ability to think.

AI Is Social Media Plus an Alibi

There is an obvious objection here, and it is worth taking seriously: this sounds like the social media discourse. We have all read the smartphone is the modern hypodermic needle essays. Why is AI coding different from doomscrolling?

It is different in one specific way, and the difference is what makes it worse.

Social media is consumption. The scroll is pure intake; nothing is produced. That is why the discourse around it eventually converged on shame as a regulator — most people know, at some level, that an hour spent on TikTok was an hour spent on nothing, and the knowledge produces a friction that limits use. The mechanism is dopaminergic but the alibi is thin.

AI-assisted coding is consumption disguised as production. Every cycle yields an artifact: working code, a closed ticket, a shipped feature, a passing test. The artifact provides legitimization the scroll cannot. You cannot tell yourself you should stop, because every prompt has output and every output justifies the next prompt. The guilt that limits social media use is absent. In its place is the opposite — a quiet pride in throughput, reinforced by Slack messages and PR counts and the dashboards your company has now started rewarding.

There is also a measurement problem. A METR study released last year⁹ handed experienced open-source developers AI tooling and measured their actual productivity. They were nineteen percent slower than the control group. They believed they were twenty percent faster. The gap between perception and reality was forty-three points. This is not a story about people being wrong about details. This is a story about people being precisely upside-down about whether the tool is helping them at all.

The BCG study found a related cliff. Productivity rises as workers add a second AI tool to their workflow, rises again with a third — then drops at four. The peak is at three simultaneous tools. Past that, more tools means less output. Yet the felt experience of using more tools is the experience of doing more work. The data and the feeling diverge.

Social media degrades attention. AI coding degrades attention and produces code at the same time. The output is the problem, not the solution. It is what makes the loop legible to your manager, your future self, your bank account. The legibility is what keeps you in the seat.

Recovery Is Bounded, But Undated

There is one piece of optimistic news in Lembke's model, and it is worth not overstating. The seesaw recalibrates. If the stimulus is removed, the pain side stops accumulating, the pleasure side recovers its responsiveness, the baseline drifts back up. This is the same hedonic adaptation that lets us get used to anything — good or bad. It is bounded. It does end.

Lembke offers thirty days as a clinical rule of thumb. The number is a heuristic, not a study finding, and it is calibrated for substances people can fully abstain from. For something like AI coding — which most knowledge workers cannot quit and have no reason to want to quit — the relevant intervention is not a thirty-day fast. It is the daily restoration of neutral. A morning at natural baseline before any agentic loop is opened. A walk without the phone. A meeting in which the agentic mode is collectively switched off. A weekend.

The seesaw does not need long abstinences. It needs intervals — small, regular returns to zero, often enough that the pain side does not accumulate. The mechanism is bounded by frequency, not by duration. This is not a counsel of restraint. It is a description of how a working system stays calibrated.

Two Layers

There are two ways this argument lands on a reader, and they are different.

For the broad middle of knowledge workers — the eighteen percent of engineers, the twenty-six percent of marketers, the colleagues quietly closing their laptops at 7pm with a headache — the BCG numbers are the relevant fact. They are not having a personal failing. They are running into the limits of an attentional architecture that was built for one tool at a time, and the productivity discourse has not caught up.

For a smaller group — the builders, the founders, the engineers with three side projects in different repos — the trap is structurally harder. The same trait that makes them valuable in this field, an appetite for shipping things, is the trait the loop hooks. Every prompt is a small completion. Every completion feels like progress on the project that matters. The dopaminergic mechanism rewards the trait the field selects for. There is no clean way to oppose the two.

A small diagnostic, written from inside this group rather than outside it: the next time you sit down to work on a project that matters, ask whether the next genuinely useful action involves writing code or designing an interface. If the answer is yes, build. If the answer is talk to a customer, send an email, follow up on a warm lead, validate an assumption — and you find yourself opening a worktree instead — that is the loop talking, not the strategy. The build is the work when the work is build. When the work is something else, the build is avoidance with extra steps.

This is the part that is hardest to write honestly. The instinct to build is real, and good, and the engine behind most things worth shipping. It is not a flaw to be extinguished. It is a tool that has, for the first time in the history of software, lost its friction — and the friction was what kept it pointed at the right targets.

Coda

There is no clean answer in this piece. The next essay in this series will try for a more practical reading — what discipline looks like when you cannot quit the thing causing the problem. For now it is enough to name the loop. To notice the seesaw. To recognize the senior engineering manager's static, the worktrees fanning out, the moment when the tools start managing you.

In Vieni avanti cretino — the button-lined office, the polite manager, the red light and the green light and the telex and the whistle — the protagonist keeps saying yes to every small additional task. The collapse is the cumulative weight of all those yeses. The point of this essay is that there is still a manager asking. He is just no longer a person. And he is very, very good at being polite.