Section 7: Agent Workflows and Multi-Agent Systems
Single agents hit limits. As the scope of what an agent handles grows, the system prompt gets longer, the tool set gets larger, and the LLM starts making worse decisions because it is trying to do too many things at once. Multi-agent systems solve this by splitting responsibilities across specialized agents that collaborate to handle complex tasks.
This section covers when to split a single agent into multiple agents, the major coordination patterns (supervisor-worker, orchestrator, peer-to-peer), how to design agent roles with clear boundaries, and the common mistakes that make multi-agent systems worse than the single agent they replaced.
What We’ll Cover
Connection to Previous Sections
In Section 3, we covered execution patterns for a single agent: ReAct, prompt chaining, router, reflection. This section scales those patterns across multiple agents. The supervisor-worker pattern is a router (Section 3) where each route is a separate agent instead of a separate prompt. The orchestrator pattern is prompt chaining (Section 3) where each step is a separate agent instead of a separate LLM call. If Section 3 taught you how one agent thinks, this section teaches you how multiple agents collaborate.
Signs Your Single Agent Has Outgrown Its Design