Async Multi-Agent Orchestration with Claude

Why it matters

Orchestration mechanics decide who controls the agent graph and how results return to it. This cookbook strips away the domain task so the mechanics alone are visible.

What to practise

Build a message hub with an inbox per agent, and the send_message and wait_for_message tools. Write one tool-use loop that appends each agent’s drained inbox to its last tool result, so no agent polls. Run a fixed team of a lead and two helpers. Then give the lead create_subagents, get_status and kill_subagents. The spawn returns at once, the helpers run concurrently, and the lead collects their reports and dismisses them. Read the trace of tool calls to follow each message.

How it fits

The fixed team and the dynamic spawn are two topologies to compare in 1.4.2. The status and kill tools are the levers 1.4.4 needs: wait_for_message wakes after a 60-second timeout instead of blocking, so decide what the lead does with a helper that get_status shows has never reported, and what it accepts when one fails.