Approval queue
When an agent reaches a tool set to Approve, the run pauses and the call waits for a human. This page covers where those pauses surface, who can act on them, exactly what the reviewer sees, and the clock that is running while they decide.
Where approvals surface
Approvals in the sidebar lists every pending item in the organization — each card shows the tool, the agent that asked, the request payload, and an Open session link. The queue is a window, not a control panel: the approve and reject buttons live where the context is.
- The run page. A Task Agent Run in Awaiting approval shows a Waiting on you panel with the call’s inputs and Approve / Reject buttons.
- The session view. In a live session, the tool call block expands in place with the same choice.
Everyone in the organization can see the queue. Only the owner of the session behind an approval can resolve it — there is no reviewer assignment or routing. If broader coverage than the owner is needed, grant watchers the Agent Reviewer role so they can open any run and see what is waiting, and reach the owner.
What the reviewer sees
An approval is the exact tool call, held before execution: the tool’s name, the server it belongs to, and the full input the agent constructed. Nothing executes until the decision. Approve releases the call to run when the session resumes; Reject records the refusal (with an optional reason) and the agent continues with that answer — a rejection consumes one of the step’s attempts rather than silently ending the run.
The clock
A pending tool approval expires after one hour. Expiry fails the run — its terminal reason reads Approval expired without resolution — it is not skipped or retried. The next scheduled slot is the next attempt.
Two other timers matter around approvals:
| Timer | Effect |
|---|---|
| 1 hour | Pending tool approval expires; the run fails. |
| 7 days | A run still sitting in Awaiting approval (for example, one paused on user input rather than a tool approval) is abandoned by the scheduler. |
Elicitations — an MCP server asking the user for structured input rather than permission — pend until answered and never expire.
Approvals leave a trail
Every decision writes a signed receipt and an audit event: who approved or rejected, when, and the exact call it covered. Executions that ran under an approval carry an Approved badge in the audit log, so “which writes had a human in the loop” is a filter, not an investigation.
There are currently no push notifications — no email or chat message when an approval is waiting. Owners of scheduled agents should treat the Approvals page as part of their working rhythm, and schedules should fire when their owner is around to answer. See Designing reliable agents.