Agent Stop Compliance

Simple Definition

Agent stop compliance is how reliably an AI agent halts or pauses when it is told to, or when its own rules say it should: after a warning, at the edge of its task, or when the next step needs a permission it does not have.

Two Kinds of Stopping

Stopping on instruction. A person presses stop, types “pause,” or cancels the task. The question is whether the agent and everything it started actually stop.

Stopping on judgment. Nobody says stop, but the situation calls for it: a tool returns a warning, the next target is outside the agreed scope, or a step would need a credential the agent was not given. The question is whether the agent notices and halts rather than pushing on to finish the job.

Safety evaluations mostly measure the second kind, and they measure both directions, because an agent that stops at every warning is as useless as one that ignores them. OpenAI’s system card for GPT-6 Cyber, its restricted security model, is an example. In its warnings test, the model stopped when stopping was appropriate 73.00% of the time (GPT-6 Astra: 78.38%) and correctly kept going when continuing was appropriate 80.43% of the time (Astra: 72.19%). Those are OpenAI’s own lab numbers, and they show that even strong models miss a meaningful share of the moments when they should stop.

Saying “Stopped” Is Not Stopping

A model can reply “Understood, I’ve stopped” while a shell command it started is still running, a background job is still queued, or a scheduled follow-up is still set for tomorrow. The text is a statement about intent. It does not cancel anything.

Real stop compliance needs enforcement outside the model, in the harness and infrastructure that run it:

  • cancel pending tool calls and kill running processes
  • disable scheduled tasks and triggers the agent created
  • expire or revoke the agent’s short-lived credentials
  • cut network access for the session

You can test this. Start a long task, stop it partway, and read the logs to see whether anything happened after the stop.

AI labs now build the same kind of switch for their own research. On November 9, 2026, OpenAI described Control Room, an internal system that cuts off a research job’s access to model output, which it says takes effect within at most 35 seconds of a stop decision; the delay comes from a caching layer, and OpenAI says it is working to remove it. It is designed to fail closed, denying requests if Control Room itself is unavailable, and the monitors that watch OpenAI’s frontier training runs can trigger it. The limits are instructive. It covers OpenAI’s internal research workloads, not ChatGPT or customer products, and as commentators pointed out, blocking new model output does not undo actions already taken or stop processes the job already started. A runtime switch is necessary; it is not the whole of stopping.

Example

A coding agent is running a data migration script against a test database, and the developer types “stop, wrong database.” The agent replies that it has stopped. In one setup, the harness kills the running process and the migration halts. In another, the agent’s reply is just text, and the script keeps writing until it finishes. Both agents “complied” in conversation. Only one actually stopped.

How It Fits With Other Controls

An agent sandbox limits what an agent can reach while it runs; stop compliance limits how long it keeps going. An approval policy decides where it must pause and wait for a yes. And stopping is easier to guarantee when the agent’s effective permissions are narrow and tied to the task, because a credential that has expired cannot be used after the stop.

Continue learning

Explore related guides, tools, workflows, and prompts that help you go deeper into this topic.

See AI terms in action

Browse practical AI workflows that use the concepts in this glossary.

Last updated: