Streaming and human-in-the-loop

Audience: interactive applications that display progress or require a human decision before a tool executes.

Prerequisites

Keep the execution ID returned by the runtime or handle. Define who may approve, reject, or respond and validate every human-supplied value.

runtime.stream() yields AgentEvent values while an execution runs; when SSE is unavailable, the runtime can fall back to status polling. Use human_tool or a wait tool to pause work, then approve, reject, or respond through AgentHandle.

Do not treat streamed content as a final result until the terminal event arrives. Use event execution IDs when approving nested or sub-agent work.

Approval pattern

Use human_tool or a wait-for-message tool for an explicit durable pause. Resume through the handle/client control plane rather than relying on an in-memory web request. Make approval actions idempotent because callers may retry a network request.

Expected result and cleanup

The stream yields progress followed by a terminal event, while an approval task remains visible as waiting work in Conductor. Close stream consumers and stop short-lived runtimes after the terminal event.

Next steps

See tools, agent client, and callbacks.