Human review

Put explicit approval and judgment boundaries around consequential work.

Human review is a policy decision, not a separate execution mode. The same tool may run automatically for routine cases and require approval when its value, risk, uncertainty, or reversibility crosses a boundary.

Review by consequence

Require review for actions that are expensive or difficult to reverse: moving money, contacting a customer, changing access, deleting data, or deploying a material behavior change. Avoid making people approve routine work that the system can safely verify.

A complete review

  1. Explain the decision

    State what will happen, why it is proposed, and who requested it.

  2. Show the evidence

    Include the records, policy, history, and uncertainty that matter.

  3. Offer bounded choices

    Make the allowed actions and their consequences explicit.

  4. Record the result

    Attribute the decision and write the resulting state and event.

Do not hide policy in UI

Apps and conversations may present a review, but they do not own its rules. Keeping policy in the backend ensures people, native agents, external agents, and apps encounter the same boundary.