Human Approvals for AI Agents Without Fatigue

Give the approver the target, arguments, impact and relevant limits. Bind the approval to that action and invalidate it if the request changes. Make expiry and rejection behaviour explicit.

Reserve approval for decisions that need human judgement under the client’s policy. Repeated low-information prompts encourage automatic acceptance. Test altered arguments, reused approvals and unavailable approvers.

Use Security Review Readiness Checklist to record gaps and owners before a scoped assessment.

• The proposed action, acting identity and exact target.
• Parameters that affect impact, including amount, destination, scope and planned timing.
• The change from the current state and a recovery route where appropriate.
• An expiry time and a record of who approved which version of the request.

An approval for a configuration change to one device group should fail if the agent substitutes another group. Test that mismatch alongside an ordinary approved request. Group routine low-impact work only where a bounded policy permits it; consequential exceptions need enough context for a meaningful decision.