Skip to main content
A session can pause mid-turn and wait for a person in two ways. The wait survives worker restarts, and the answer resumes the same turn.

​
Tool approvals

Each connected account has Allow, Ask first, and Block choices. Allow runs without asking, Ask first waits for review, and Block prevents execution. Tool recommendations apply until you choose otherwise. These choices do not grant account access or additional provider permissions. For Ask first, the session enters requires_action. The review shows the account, action, consequences and selection count, with full saved details available separately. Approving resumes that exact action; declining lets the agent adapt. A waiting review retains its original choice. Changed settings apply when work next starts or resumes. Only humans can approve tool calls. Programmatic calls also retain their exact arguments through the wait. Approval resumes the stored operation; it does not restart the earlier program. An unknown outcome means execution may have started and needs reconciliation before retrying.

​
Structured human input

Agents can ask questions with the built-in request_human_input tool. A request holds up to 20 free-text, single-select, or multi-select questions, and can allow Skip and set a deadline. The agent receives one of these outcomes: Workspace admins can turn the tool off with the agentHumanInputEnabled setting.

​
Sending a message while the session waits

An ordinary message sent during either wait cancels it and is delivered as a Steer, so “actually, do this instead” is one action.

​
Answering from your product

SessionConversation from @opengeni/react renders both kinds of wait. To answer from your own UI or backend, see Approvals & questions.