Human-in-the-Loop Approvals
Last Modified: 2026-11-19

Human-in-the-Loop Approvals

Some actions are important enough that a person should approve them before they run. A policy rule can be marked as needing approval; when an agent makes a call that matches that rule, the human-in-the-loop (HITL) queue lists the action for a human decision before it proceeds. Agent ID supports more than one elicitation mode for getting that decision from a human reviewer.

Elicitation ModeHow It Works
In-tool approval promptThe gateway pauses the call and shows an approval prompt directly in the person's AI tool (their MCP client). If they approve, the call goes ahead; if they decline, or do not respond in time, the call is blocked.
CIBA The gateway sends an out-of-band push notification to the approver's CIBA-enabled identity-provider client; the approver clicks Yes or No independently of the agent's session.
Email The approver receives an emailed link to review and approve or deny the request outside the calling session.

Note:  If the AI tool cannot show an approval prompt, the call is blocked rather than allowed. Agent ID never lets a call through just because it could not ask.

Procedure 

  1. On the Approval requests page, click the pending request.
  2. Review the requesting identity, the tool and resource it is calling, and any supporting context shown.

Results 

The originating agent's call completes if approved, or is rejected if denied, and the decision is written to the audit log.