Approvals & ask-user
Pause for approval or clarification, then resume
Human-in-the-loop
The execution engine can pause in two ways: when a sensitive endpoint requires approval, or when the AI decides it needs more information from you. In both cases, execution stops, stores a resume state, and continues only after you respond.
Endpoint approval
Some agent endpoints can be marked as requiring approval before they run (e.g. production deploys, destructive actions). When the executor chooses such an endpoint, it does not run it immediately. Instead it emits an approval required event with the planned task and inputs. The UI shows a prompt for you to approve or reject. If you approve, execution resumes and the task runs; the rest of the loop continues as normal.
Resume state
Your approval (or rejection) is sent back with the stored resume state. The executor continues from the same iteration with your decision; no need to re-describe the goal.
Ask user
The AI can decide that it needs clarification or a choice from you (e.g. "Which environment: staging or production?"). It emits an ask_user event with a question and optional options. Execution pauses until you reply in the chat. Your reply is fed back into the next Think step, and the executor continues with the new context.
Resume state
Resume state is a snapshot of the executor at the pause point: iteration count, tasks already executed, their outputs, and the pending approval or question. It is stored with the conversation so that when you respond, the same execution can resume without losing progress.
| Concept | Description |
|---|---|
| Resume state | Snapshot of executor state used to continue after user input |
| Approval required | Endpoint is configured to require approval before execution |
| Ask user | AI requests clarification or a choice; execution pauses until reply |