A good agent runbook does not only say what to do next. It also says when to stop. Stop conditions are the small rules that keep automation from turning uncertainty into momentum.

The most useful stop conditions are simple enough to follow under pressure. They do not require the agent to infer intent, guess acceptable risk, or decide whether a surprising state is "probably fine." They name the exact signal that ends autonomous action.

Stop when evidence stops matching the plan

Agent work often begins with a reasonable plan and then discovers a different system than expected. That is not failure; it is new evidence. The unsafe move is to keep executing the original plan after the evidence has changed.

Useful stop conditions include:

  • the target file, route, service, or record is missing or duplicated
  • the observed state conflicts with the runbook's assumptions
  • a validation command fails for a reason unrelated to the intended change
  • the agent cannot explain what will be different after the next action
  • the next step would expose private data, credentials, identifiers, or logs

These rules are intentionally plain. They give the agent permission to stop before it has to improvise.

Put halt rules next to action rules

Stop conditions work best when they appear directly beside the step they constrain:

Regenerate public indexes after content changes.
Stop if the generator reports a schema error, changes unrelated generated assets, or produces output that cannot be explained from the content edit.

That pairing prevents a common runbook bug: the action is specific, but the safety boundary is somewhere else or missing entirely. If a step can change state, publish, delete, notify, rotate, or overwrite, it deserves a local halt rule.

Report the stop as useful output

Stopping is not the same as disappearing. When an agent hits a stop condition, the final report should include the goal, the stop signal, the last safe evidence, and the smallest suggested next decision. That turns a halted run into a reviewable handoff instead of a vague failure.

A compact format is enough:

Stopped before deploy.
Goal: publish one public-safe note.
Stop signal: readiness check failed after index regeneration.
Last safe evidence: content file was created; generated assets were not manually edited.
Next decision: fix the validation error or revert the draft.

Stop conditions make autonomous work more dependable because they convert uncertainty into a controlled pause. The agent still helps, but it no longer has to pretend that every situation is safe to solve alone.