The stop button is a decision

We describe the kill switch as the thing that saves us from the machine. It is the reassurance offered at the end of every anxious conversation about autonomous systems: whatever else is true, a human can always pull the plug. The circuit breaker, the emergency stop, the reviewer who can freeze a run mid-stride — these are named as safety, and rightly so. But there is a strange asymmetry in how we treat them. We build agentic systems that log every action they take with care — the tool called, the argument passed, the result returned — and then we treat the moment a human reaches in and stops the whole thing as an operational event, something that happened to the system rather than something the system must account for. The stop is the one act we are most confident matters, and the one we are least likely to record as a decision.

It is a decision, though. When an agent is halted, someone or something chose to halt it — at a particular point, on a particular signal, leaving the work in a particular state. When an agent is allowed to run past a checkpoint that could have stopped it, that too was a choice, made by whoever declined to intervene or whatever rule declined to fire. Both are exercises of authority over an autonomous process. Both can be right or wrong. And both can matter more than any single action the agent took on its own — because they are the points at which the agent's autonomy was overridden or ratified by a judgment from outside it.

Stopping is deciding

Consider what is actually contained in a halt. There is a trigger: something crossed a threshold, or a person watching a dashboard decided they had seen enough. There is an actor: a human operator, an automated guardrail, a supervising system. There is a moment: the agent was three steps into a five-step process, or partway through a transaction, or between two writes that were supposed to be atomic. And there is a resulting state: what was left half-done, what was rolled back, what was frozen in place. Each of these is a fact about a consequential decision, and each is exactly the kind of fact that recollection cannot recover after the fact. Nobody remembers, three weeks later, the precise signal that prompted the intervention or the exact point in the run where the brake was pulled.

Yet these are the facts most systems drop. The reason is partly architectural and partly a failure of imagination. Architecturally, the stop usually lives outside the agent's own record-keeping — it is a control-plane event, a signal sent from the operator's console, not an action the agent itself performed and logged. The agent records what it did; it does not record what was done to it. And imaginatively, we think of the halt as the moment accountability was satisfied rather than a moment that itself requires accounting. The human intervened; surely that is the responsible outcome, not a new thing to be responsible for. But this gets the logic backward. The intervention is not the end of the accountability question. It is the sharpest instance of it.

The intervention is where responsibility concentrates

An autonomous agent, running inside its granted authority, is in some sense the least interesting part of the accountability picture — it is doing what it was built and permitted to do. The interesting moments are the seams, and the stop is the most important seam of all. It is the point where human judgment re-entered a process that had been proceeding without it. Everything upstream of the halt was the machine acting on its authority; the halt is a person, or a rule standing in for a person, asserting authority over the machine. Responsibility does not dissolve at that seam. It concentrates there.

This is why the record of the stop assigns the act in a way that little else does. If an agent is halted mid-transaction and a customer is left in a broken state, the question that matters is not only what the agent was doing but who decided to stop it there, and on what basis. If an agent is not halted — if it runs past a checkpoint and does damage that a timely intervention would have prevented — the question is who or what held the authority to stop it and did not. That non-halt is a decision with an author, even though nothing visible happened at the moment it was made. Silence, here, is not the absence of a choice; it is a choice that happens to look like inaction. A system that records only the halts and never the deliberate decisions to continue has recorded half the interventions and named the easier half.

A system that records every move an agent makes but not the moment a human stopped it has logged the machine and lost the person.

The point of recording the stop is not to second-guess the operator who pulled it. Most interventions are correct, and the ones that are correct deserve a record precisely because they demonstrate that the authority to halt was exercised well and in time. A stop with a clean record is not an accusation; it is proof that someone was watching and acted. The record protects the intervener as much as it exposes them. What it forecloses is the failure mode where authority is exercised invisibly — where a run is killed, a state is left broken, and there is no durable answer to the question of who reached in and why. Authority exercised without a record is authority no one can be held to, and that is true whether the exercise was wise or reckless.

Recording the brake

What does it take to treat the stop as a first-class decision rather than an operational blip? Not much that the system is not already capable of, which is part of what makes the omission frustrating. A record of a halt has to carry the same things any accountable decision record carries, adapted to this particular act. It needs the trigger — the signal, threshold, or human observation that prompted the stop, captured as it stood at the time rather than reconstructed later. It needs the actor — the specific person or automated authority that issued the halt, so the act has an identifiable author with standing to have made it. It needs the agent's state at the moment of intervention — where in its process it was, what it had completed, what was left pending or partial. And it needs the downstream effects — what the halt left behind, what was rolled back or reversed, what remained in an inconsistent state and had to be reconciled by hand.

With those four things, the stop stops being a gap in the account and becomes part of it. The difference between an operational blip and an accountable intervention is entirely a difference in what was captured. A halt that leaves only a line in a system log saying the process was terminated tells you nothing about the decision — it records that the machine stopped, not that a person decided to stop it, on what, in what state. A halt captured as a decision, with its trigger and actor and state and consequences, can be examined the way any consequential judgment can be examined: contested if it was wrong, defended if it was right, reconstructed by anyone with standing rather than narrated from memory by the one person who happened to be at the console. The same is true, exactly, of the decision not to halt — a recorded checkpoint that fired or declined to fire is an account of a judgment; an unremarked continuation is a judgment that left no trace.

The frontier of agentic autonomy is often framed as a question of how much we should let these systems do on their own. That is the wrong axis. The systems will act autonomously; the question that decides whether the autonomy is answerable is whether the moments when a human reached in — or pointedly did not — are captured as the decisions they are. The kill switch is not outside the accountability system. It is the most important thing inside it. A halt without a record is the one intervention that mattered most, exercised by an authority no one can name, on grounds no one preserved. If we are going to keep telling ourselves that the human can always stop the machine, we owe ourselves a record of every time the human did — and every time the human could have, and chose not to.

— Dispatches · Summit Cognitive