Skip to content
Canon

Volume IV

Multi-Agent Architectures

Crews, graphs, swarms, and societies

A second agent is a context boundary, not a personality. The useful topologies are few: supervisor, crew, conversational society, handoff swarm, debate. The failure mode is always the same — shared, bloated context and unbounded talk.

  1. IV.1Orchestrator–WorkersA central model decomposes work it cannot predict in advance, delegates to workers, and synthesizes. The multi-agent pattern Anthropic actually endorses.7 min
  2. IV.2Role-Playing CrewsCrewAI's bet: give each agent a role, a goal, and a backstory, assign tasks, run a process. It is the fastest way to sketch a team — and the easiest way to overfit to theater.6 min
  3. IV.3Conversational Multi-AgentAutoGen's original idea: agents as speakers in a group chat, with a manager selecting the next speaker. Powerful for research; dangerous as a production bus.6 min
  4. IV.4Handoffs and SwarmControl transfers. The current agent returns the next agent. Context is a clean brief, not a pile of speakers. OpenAI Swarm made this obvious; the Agents SDK made it shippable.6 min
  5. IV.5Hierarchical SupervisorA supervisor node routes to worker nodes and back until a completion condition. LangGraph's most copied multi-agent topology — and the one that actually checkpoints.6 min
  6. IV.6Society of MindMetaGPT's software company: SOPs as prompts, specialized roles producing artifacts for the next role. Minsky as a Makefile.6 min
  7. IV.7Debate and CritiqueTwo or more models argue; a judge or a rubric decides. Use when single-pass generation is biased and you can afford the tokens.5 min