How to review and approve agent actions
Sensitive tool calls pause mid-run for an explicit, recorded human decision — rolling out with the V1 Beta.
Every tool an agent declares carries a policy: allow, ask, or deny. When a tool is set to ask, the run pauses at the moment the agent wants to use it and waits for a human decision.
Beta
Human-in-the-loop approvals are rolling out with the V1 Beta. The runtime today executes allow-listed tools only and refuses everything else.
Approving in chat
In a conversational session, the request appears as an inline confirmation card: what the agent wants to do, with which inputs. Approve to let the action run, or deny — optionally with a reason. The session continues either way.
Approving in task mode
For task runs, the calling product surfaces the approval request to the right user, and the answer is sent back through the run's approval endpoint. The run resumes as soon as the decision arrives.
Timeouts and audit
Pending approvals time out after 30 minutes by default, and a timed-out request is treated as denied. Every decision — who approved or denied, when, and why — is recorded as a first-class entry in the audit log.
Still stuck?
Tell us what you were trying to do and where it failed — screenshots and run IDs help. A human follows up.