Skill · pattern · multi-agent
Multi-agent topologies
Choose supervisor, crew+flows, or handoff swarm. Use when adding a second agent, CrewAI, AutoGen group chat, 'team of agents', society of mind, or role-play. A role is real only if tools or schema change. Never unsupervised group chat on production tools.
GET /api/canon/skills/multi-agent?format=md
The useful topologies are few. The failure mode is always shared bloated context and unbounded talk. CrewAI added Flows for a reason. AutoGen group chat is a research primitive, not a prod bus.
When
Specialists need different tools or a fresh window, and a parent must synthesize.
Do
- 01
Pick one topology
Supervisor in a durable graph (LangGraph/Mastra) for checkpoints. Crew+Flows when jobs are nameable and artifacts pass. Handoff swarm when a specialist should own the conversation.
- 02
Roles must change the ACI
Different tools, prompt constraints, or schema. Backstory without that is theater.
- 03
Bound the talk
Typed messages, turn caps, no prod tools on the chat. Prefer supervisor or handoff over group chat.
Don't
- Unsupervised group chat against a production database.
- Four personas sharing one transcript.
- A Genius/Sage/Critic trio with the same function-calling surface.
Hard rules
- Second agent = context boundary.
- Fresh or sliced windows. Structured returns.
- Society-of-mind / debate only with a typed protocol and a cap — never as the outer loop on tools.
Refuse
- Shared mega-context — A crew sharing one bloated window. Specialists drown in each other's work.
- Role-play without a different ACI — Backstories that do not change tools, prompt constraints, or schema. Theater.
- Unsupervised group chat — Agents talking to each other against production tools with no turn bound and no typed messages.
- Handoff history dump — Transfer of control copies the entire conversation into the specialist.
Load with this
Load next
- Spawn subagentsProtect the parent window, parallelize, or apply a different skill set.
- Orchestrator–workersSubtasks are not known up front. The parent must not drown in search hits.
- Router / handoffStable categories and a cheap, accurate classifier.
- Workflows versus agentsYou are about to pick a topology. Most production systems that work are workflows with a small agentic island.
Trigger tests
Should fire
- “Build a multi-agent crew”
- “Should we use AutoGen group chat?”
- “PM / engineer / QA agents”
Should not
- “A single coding agent on one repo”