Why it matters
Single agents get overwhelmed by complex tasks. Multi-agent decomposition improves quality, testability, and modularity. Understanding the pattern is essential for anything beyond trivial agents.
The architecture
Coordinator pattern: top agent has tools that invoke sub-agents. Sub-agents are themselves agents with their own tools and prompts, focused on specific domains.
Specialists: search agent has web tools, code agent has execution tools, data agent has database tools.
How it works end to end
Delegation flow: coordinator plans, invokes sub-agent as a tool with a specific sub-task, receives sub-agent's output, integrates into overall response.
Communication: keep sub-agent interfaces narrow (well-defined input/output). Don't expose full conversation to every sub-agent.
Cost: multi-agent uses more tokens (each agent has its own context). Weigh against quality gains.