Human-in-the-Loop Agents

Human review should appear where risk or ambiguity is high, not as a blanket prompt after every harmless lookup.

Risk-based approvals

Require review for destructive changes, external communication, financial actions, production deployment, or access to especially sensitive data.

Pause and resume

A robust runtime serializes enough state to resume after minutes or days. OpenAI's Agents SDK models approvals as interruptions and supports serializing run state.

Make the approval meaningful

Show the reviewer the exact action, arguments, target system, and consequences. “Allow tool?” without context is weak oversight.