A banking multi-agent system can propose wire transfers. Policy requires human approval for wires over $10,000, supports manager override on edge cases, and must record who approved what. How should you implement human-in-the-loop?
Select an answer to reveal the explanation.
Short Explanation
You need a real human on the critical path before the risky tool runs. Pause at a durable approval step, capture who approved and what they decided, allow reject or override paths, then execute the payment only after approval.
Full Explanation
HITL includes approval workflows, overrides, and edge-case handling. Pausing before the wire tool at a durable approval step, capturing approver identity and decision, supporting reject/override paths, and executing only after approval creates a true control point with auditability before side effects.
Sending the wire immediately and emailing afterward inverts control and fails preventive policy. Model role-play as manager is simulated oversight, not compliance-grade human control. Optional comments that do not affect tool order decouple governance from enforcement.
Use durable orchestration (queues, workflow engines, or Agent Framework human approval patterns) so waits survive restarts. Exam tip: approvals must gate tool calls with irreversible external effects.