How to wire a multi-agent crew

Split work across an orchestrator and bounded workers.

1. Define one orchestrator with the user goal and worker roster

2. Give each worker its own tool allowlist and step cap (e.g. max 3)

3. Workers return structured JSON summaries, not raw tool dumps

4. Orchestrator merges summaries into the final answer

5. Enforce a **global step cap** across all agents to prevent runaway loops

**Try the lesson:** `multi-agent-in-code` in Lane D