The approval step sits between an agent's proposal and a tool that changes external state. The reviewer should see the exact target, effect, evidence and recovery information rather than a vague confirmation request.
Approval is most useful for actions with material or difficult-to-reverse consequences. Routine low-risk analysis can remain automated while deletion, sending and production changes receive deliberate review.
ELI5
Human-in-the-loop command approval means an AI system can prepare an important action but must wait for a person to approve the exact action before it happens. This separates drafting and analysis from execution in another system.
For example, an agent might draft an email and show the recipient, subject, and complete message, but it cannot press send until the user agrees. The approval should cover only what was shown and must not silently authorize a different recipient or later action.
