How to Add Human Approval to an AI Agent

Add human approval at the boundary where an AI agent is about to perform a consequential action. Register the agent with the minimum required scope, create an approval or Question containing the proposed action, and keep the agent paused until PingRoom returns an explicit allowed outcome.

Make the request reviewable

Identify the action, target, consequences, and relevant resource version. Deploy revision 8f31c2 to production is reviewable; Continue? is not. Attach a safe link to evidence when the reviewer needs more detail, but never send credentials or sensitive prompt content in the notification.

Use a stable correlation ID so the decision can be joined to the same agent run. Set an expiry appropriate to the action. Treat denial, cancellation, expiry, network failure, and no response as distinct outcomes—none of them is approval.

Enforce the result outside the notification

The tool or service that performs the action must verify authorization and consume the decision once. PingRoom is the human decision channel, not a substitute for deployment permissions, database controls, or cloud policy.

Keep the agent credential separate from the human account token. Limit it to the rooms and tools the agent actually needs, rotate it when exposed, and record the final action in the owning system’s audit log.

For low-risk informational events, send a plain structured Ping instead of manufacturing an approval. Human attention is a scarce control surface.

See the agent access documentation, CI approval example, and Question Protocol for the complete pattern.