
Multi-Agent Systems: When More Agents Help or Hurt
Michael Ryan, a Stanford CS PhD student and contributor to the DSPy project, argues that multi-agent AI systems are not automatically better than a single capable agent. Teaching session 8 of Stanford's CS329Z: Engineering AI Agents, Ryan lays out when to split work across multiple agents versus keeping one, and walks through orchestration patterns including orchestrator-worker, swarm, and hybrid designs. He covers how agents hand off tasks and communicate through message boards and Google's Agent2Agent protocol, and the coordination tools that keep them from stepping on each other, including task ledgers, shared memory, and git worktrees. The back half surveys published research on why multi-agent systems fail in practice, from coordination overhead to compounding errors. The lecture is aimed at students building agentic AI systems and assumes familiarity with the course's earlier sessions on single-agent design.