Good operators plan for correction before a task goes wrong.
If the plan is wrong
Stop before more work starts. State which assumption or domain is wrong, update the context and ask Inspector to propose a corrected plan. Do not preserve a bad decomposition simply because agents have already started.
If evidence is weak
Ask for the exact source, a stronger source class or an explicit uncertainty label. If the missing evidence is decision-critical, the correct result may be "not enough evidence yet".
If an action is wrong
- Stop additional writes.
- Record the exact action and target.
- Verify whether the action completed or partially completed.
- Use the documented rollback when one exists.
- Escalate to the system owner if reversal is uncertain.
- Update the operating constraint before continuing.
Reusing successful work
Repeatability comes from a stable decision brief, known sources, a bounded plan and a consistent review rubric. Keep those as a reusable operating template. Do not assume the same sub-agent team is right next month if the decision or evidence changed.
For a recurring review, name:
- the human owner,
- the trigger or cadence,
- the current source set,
- the approval boundary,
- the output location,
- the failure notification,
- the stop condition.
You now know the real Inspector control loop: discover, plan, approve, delegate and synthesise. You also know how threads retain context, why dynamic agents replace a fixed roster, how broad tool access changes operator responsibility, and how to recover when evidence or actions are wrong.