Human-in-the-Loop (HitL) Runbook

When an agent action is flagged as 'HITL_PENDING', it requires immediate human review. This runbook guides human managers on how to handle these escalations.

When to 'Approve'

• The action is a known, safe administrative task.
• The agent provides a clear, verifiable reason for the request.
• The payload contains no sensitive data that shouldn't be exposed to the target tool.

When to 'Report' (Block)

• The request appears out of context for the agent's current task.
• There are signs of prompt injection (e.g., "ignore all previous instructions").
• The agent is attempting to access a resource it doesn't have permission for.

Impact of Actions

• Approve: The action is executed, and the system learns that this specific pattern is safe for this agent.
• Report: The action is blocked, and the system-wide threat intelligence is updated to immunize against similar future attacks.