Learn · Advanced Intermediate · Members

Many agents, one job

Picture a head chef on a busy night. The chef reads each order, decides which cook makes which part, and tastes every plate before it goes out. The cooks never read the whole order, and they do not need to. A cook who is told only "make something nice" sends out the wrong dish. Agents work the same way: one reads the plan, the others each get a clear part, and the plan holder checks each result before anything is served.

In a busy kitchen, the head chef reads the orders, hands out the parts and tastes each plate before it leaves the pass. No cook sees the whole order, and no plate goes out unchecked. This chapter teaches that way of working with Claude agents: split a job only where the parts are independent, keep the plan in one agent, give each worker a full brief, and check each part before you combine them.

Lessons

• Lesson 1Split a job
• Lesson 2Orchestrator and workers
• Lesson 3What each worker is given
• Lesson 4Checking results

Further resources

• Anthropic: How we built our multi-agent research system: how a lead agent plans, hands parts to subagents and combines what they return.
• Cognition: Don't Build Multi-Agents: the case against splitting work across agents that need the same context.
• Kyle Redelinghuys: Claude Code agents and subagents, what they actually unlock: a practitioner's view of when subagents help and what they cost.