Multiagent: watch a curriculum team work in real time

Why it matters

A coordinator’s session stream shows only its own thread, so a subagent can work for minutes with nothing visible. 3.4 asks what a multi-agent run lets you observe; this notebook closes that gap on one platform.

What to practise

The notebook runs on Claude Managed Agents, a beta API. Create a coordinator with two specialists, setting model.effort to high on the researcher only, and start the session with initial_events on sessions.create. Read the session-level stream, and on each session.thread_created event attach a watcher to that thread’s own stream with event_deltas, exiting on session.thread_status_idle. Deltas are best-effort and never replayed, so the preview is a scratch buffer and the buffered agent.message is the record. Compare each preview with its record. Then update the researcher without passing version, and note that the coordinator keeps delegating to the version its roster pinned until it is updated too.

How it fits

The per-thread streams answer 3.4.1’s observation gap for multi-agent execution, and each thread is a subagent segment of the kind 3.4.2 traces. Compare the preview-versus-record rule with 3.4.1’s split between execution and outcome: diagnose from the buffered record, never from the preview.