The Judgment of When to Stop

Stopping is a learned skill, not a personality trait. An agent, left to itself, continues — the next paragraph, the next test, the next variant, indefinitely. Only a human knows what "enough" means, because "enough" is a judgment against a standard the human carries and the agent does not. In the era of cheap continuation, the stopping decision is the densest cognitive act a knowledge worker performs.

This essay is for anyone who has watched an agent produce a fourth, fifth, sixth version of something and felt the strange new difficulty of the moment: not making the work better, but declaring it done.

For most of the history of work, nobody needed to be good at stopping, because the world stopped you. The body priced continuation — fatigue arrived, attention frayed, the light failed. Deadlines priced it. Budgets priced it. Every additional hour of polish cost an hour of something else, and the cost was felt. Stopping was less a decision than a collision with a limit, and because the limits were external and reliable, the judgment of when the work is enough never had to be developed as a skill in its own right. The environment supplied it.

The first essay in this series traced what happened when the cost of generation collapsed: value migrated to taste, the trained capacity to perceive which output is worth keeping. This essay is about taste's companion skill, and it begins where the cost collapse bites hardest. When continuation becomes nearly free — when the next draft, the next variant, the next hour of machine polish costs almost nothing and arrives in seconds — every external stop signal goes quiet at once. No fatigue, because the agent does not tire. No marginal cost worth noticing. No natural pause in which the question is this done? used to force itself. What remains is the internal signal. And the internal signal, it turns out, is exactly what seventy years of decision research has been describing.

#Why is stopping suddenly the hard part?

Because the default flipped.

An agent's operating loop has no internal stop. It will extend the essay, refactor the module, generate the alternative — not because it is ambitious but because continuation is its resting state. Whatever halts it is external: a token limit, an interruption, or a condition that a human wrote into the specification. Even the most sophisticated agent stop-condition is a human stopping judgment, exported in advance into text. The judgment did not leave the system. It moved upstream, onto the person.

This inverts the historical arrangement. Work used to be effort-bound: producing more was expensive, so the scarce act was continuing. Now producing more is trivially cheap, so the scarce act is ending. The person who cannot stop does not fail loudly in this new regime — they fail quietly, in the specific way the regime makes available: another round of variants, another pass of polish, another regenerated candidate, while the thing that was already good enough to ship sits unshipped. The failure mode of the effort-bound era was abandonment. The failure mode of the agent era is endlessness.

The productivity literature has spent decades on how to start — rituals of beginning, activation energy, the first five minutes. Almost nothing in it treats ending as a skill. That omission was affordable when the environment did the ending for us. It is not affordable now.

#What did Herbert Simon actually discover?

In 1956, Herbert Simon published a short paper in Psychological Review with an unassuming title — "Rational Choice and the Structure of the Environment" — that contained one of the most consequential ideas in twentieth-century social science.¹ The reigning model of rational choice held that a rational actor surveys the alternatives, evaluates each, and selects the best. Simon's observation was that no real organism ever does this, because no real organism can. The alternatives are too many, the evaluation too expensive, the time too short. What organisms actually do, he argued, is satisfice: they search until an option clears an internal threshold of acceptable — the aspiration level — and then they stop.

The crucial part is what satisficing is not. It is not a compromise, and it is not laziness dressed up in theory. Simon's argument, extended across the work that earned him the 1978 Nobel Prize in economics, is that satisficing is the correct strategy for any decision-maker with bounded time and bounded attention — which is every decision-maker that exists.² Optimization is a fiction that assumes free search. Under real constraints, the intelligent move is to hold a threshold and stop when the world clears it. The threshold adjusts with experience: too many options clear it too easily and it rises; nothing clears it for too long and it falls. The mechanism is adaptive, and it is the closest thing decision science has to a definition of practical wisdom.

Half a century later, Barry Schwartz and colleagues measured what happens to people who refuse the satisficing move — the maximizers, who keep searching past acceptable for the best. Across their studies, stronger maximizing tendencies correlated with lower reported happiness, lower life satisfaction, more depression, and markedly more regret.³ The correlational design means the arrow of causation stays officially open. But the pattern is consistent and hard to ignore: in the data, the inability to stop searching does not look like a higher standard. It looks like a burden.

#What happens in the mind at the moment of enough?

Decision science has a surprisingly specific answer.

The best-supported family of models for how a decision resolves is sequential sampling, with the drift-diffusion model at its center. Roger Ratcliff and Philip Smith's landmark comparison established the framework's dominance for simple decisions: as a person considers, evidence accumulates moment by moment toward one option or another, and the decision fires at the instant the accumulated evidence crosses a threshold.⁴ The framework has since been extended and stress-tested across two decades of research.⁵ Its deepest insight is easy to miss: the threshold is a parameter, not a constant. The decision-maker sets it — mostly without knowing they are doing so — and where they set it expresses a tradeoff. High thresholds buy accuracy at the cost of time. Low thresholds buy speed at the cost of error. Experience improves the evidence a person can extract from what is in front of them — and it also, less famously, improves where they place the line the evidence must cross.

Honesty requires the caveat: drift-diffusion was built for second-scale choices — is the dot moving left or right? — not for deciding whether an essay is finished. Nobody has run a diffusion study of manuscript revision. But the architecture the model describes is the best available account of what a stopping decision structurally is: evidence about the work accumulating against an internal criterion, and an act that fires when the criterion is crossed. And it gives the craft literature's oldest intuition a precise object. Alongside the skill of seeing the work clearly, there is a second, separable skill — knowing where to hold the line the work must cross — and thresholds sharpen the way Simon said aspiration levels do: by repeated setting, testing, and adjusting against outcomes.

This is also, precisely, what the agent lacks. An agent evaluates against the criteria it was given; it does not carry the unstated standard that the criteria were an attempt to approximate. When the output is technically compliant and somehow still wrong — every practitioner of the agent era knows the experience — what has been detected is the gap between the written criterion and the carried one. Detecting that gap is a threshold-crossing event in a person. There is no analogous event in the machine.

#Can stopping be trained?

The strongest evidence that stopping is a practice rather than a temperament comes from the people who did it deliberately for decades.

In October 1935, Hemingway published a piece in Esquire called "Monologue to the Maestro," advice to a young writer who had turned up in Key West wanting to learn the trade. Its central instruction has been quoted ever since, usually without noticing how procedural it is: "The best way is always to stop when you are going good and when you know what will happen next."⁶ Not when you are stuck. Not when you are empty. When you are going good — at the exact moment continuation feels most natural and most justified. He described the same practice again, decades later, in his Paris Review interview: work until the day's piece is done and the next move is visible, then stop and do not think about it until tomorrow.⁷

Read as decision science, the practice is a threshold rule of unusual sophistication. It ends the session on completion rather than depletion. It banks a known starting point, which dissolves tomorrow's cold start. And it hands the developing material to the hours in between, on the accurate conviction that the next move ripens off the desk. What it refuses is the maximizer's move — the seductive one-more-paragraph while the going is good — because Hemingway understood that the going-good is precisely the resource a stopping rule exists to protect. The practice held for the better part of four decades. That it eventually failed him, when the conditions it depended on could no longer be assembled, is part of the record too — we have written that portrait separately.

He was not an outlier in kind, only in fame. Csikszentmihalyi's interview studies of creative practitioners document, again and again, working lives organized around deliberate session boundaries — fixed morning hours, fixed rituals of ending — rather than around working to exhaustion.⁸ The logic of the arrangement is not hard to reconstruct: deep engagement is not sustainable indefinitely, and an ending decided in advance protects the next session from paying for this one's depletion. Ericsson's Berlin violin study — the founding dataset of the deliberate-practice literature — found a matching structure at the level of hours. The best violinists accumulated more deliberate practice than their peers, famously; less famously, they did it in concentrated sessions rarely exceeding ninety minutes, rested between them, and slept more than their merely good colleagues.⁹ The elite difference was not a refusal to stop. It included, measurably, the discipline of stopping.

Cal Newport gave the day-scale version of the practice its modern name: the shutdown ritual, a fixed sequence that reviews open commitments, queues tomorrow's first move, and closes the working day with an explicit marker.¹⁰ The mechanism it addresses is one of the better-documented effects in the attention literature: work that ends without being deliberately closed does not actually end. Sophie Leroy's attention-residue studies showed that tasks left open continue to occupy cognitive bandwidth after the switch, degrading performance on whatever comes next.¹¹ A bad stop, in other words, is not merely a missed opportunity to rest. It is an open loop that keeps billing. We have built Particle's evening around exactly this research.

Across every scale — the sentence, the session, the day — the pattern is the same. Stopping well is a rule, rehearsed until it becomes a capacity. Nobody in the record was born knowing when to stop. They built the threshold, and then they kept it.

#What we built, and why

Particle's most argued-over design decision is also its smallest: a session ends in one of two ways, and the ways are not symmetric. COMPLETE means the human declares the work done — the counter advances, the celebration fires, the particle joins the pile. SKIP means the session ended without that declaration — elapsed time is recorded, nothing is celebrated, nothing counts. The asymmetry is the entire point, and this essay is the argument for it.

The rest of the architecture follows the same research. Sessions have planned durations — the ending is the default, not the exception, which is Csikszentmihalyi's practitioners and Ericsson's bounded sessions encoded as product structure. The REFLECT stage of the Particle Loop is the calibration step: a completed session is compared, briefly and honestly, against the standard it was meant to clear, which is how a threshold gets feedback and how stopping improves with use instead of staying an arbitrary feeling. And Wind Down closes the day the way Newport prescribed, because an unclosed day leaks residue into the evening that the next morning has to repay.

None of this tells the user when their work is enough. That judgment is theirs, unstated standard and all — the same reason the Coach observes but never advises. What the product does is guarantee the judgment gets practiced: one real stopping decision per session, several per day, at a scale where a miscalibrated threshold costs twenty-five minutes rather than a quarter. The reps are small. The capacity they build is not.

#What does stopping become when continuation is free?

The load-bearing skill of the agent era.

Consider where the stopping decision now appears in a single working day of agent-assisted work. When is this prompt specified enough to send? When is this draft the one to keep, against the near-free temptation of one more regeneration? When is this conversation with the machine producing diminishing returns? When is the feature — which the agent would happily extend forever — done? Each of these is the same cognitive act: evidence against an internal threshold, resolved by a human who carries a standard the machine cannot see. The person who performs that act well ships. The person who performs it poorly becomes the bottleneck in their own pipeline, presiding over an ever-growing field of almost-finished things — and almost-finished things, as we have argued elsewhere, do not accumulate into anything.

There is one more asymmetry worth naming plainly. The old world's stopping signals were bodily — you stopped, eventually, because you were spent. The agent removes even that. It works while you sleep; it never runs down; the traditional last-resort cue is simply absent from the loop. Whatever stopping happens in the agent era happens because a human decided it should. The decision has lost its safety net, which is exactly why it has to become a trained one.

Taste, the first essay argued, is knowing what good looks like. Stopping is acting on the signal — the moment perception becomes commitment. The next essay, The Prompt as Document, turns to the skill that precedes them both in the pipeline: rendering intent into a specification an agent can execute, which is where every exported stop-condition is written. The fourth, The Architecture of Trust, treats stopping's mirror image — when to stop verifying. And the finale, The Last Instruction, closes the series with the largest stopping decision there is: knowing when to abandon the plan itself.

Simon's threshold, Hemingway's rule, Newport's ritual — the research and the practice agree, and have agreed for seventy years. Enough is not a feeling you wait for. It is a line you learn to hold.

Hold it once today. Stop while you are going good, and know what happens next.

#References