Multi-Agent Wiki

Taxonomy

A five-dimensional taxonomy of multi-agent patterns

Don't classify multi-agent patterns by framework name. Classify them by engineering dimension. Any real system usually composes several patterns at once — for example: Supervisor + Parallel Fan-out + Blackboard + Verifier + Human Approval + MCP.

1. Control structure

TypeWhere control livesRepresentative patternsWhen it fits
CentralizedOne coordinator / manager owns the flowSupervisor, Agents-as-tools, RouterStable production flows, support triage, enterprise tasks
TransferThe active agent can hand off controlHandoffMulti-domain expert switching, long conversations
HierarchicalManagers manage workers; workers may decompose furtherHierarchicalLarge projects, cross-team, multi-phase work
DecentralizedNo fixed center; agents talk or relay on their ownPeer-to-peer, SwarmOpen exploration, autonomous networks

2. Information flow

TypeHow information movesRepresentative patterns
LinearA's output feeds B, B's output feeds CSequential Pipeline
ParallelMany agents run concurrently; results mergedParallel Fan-out / Gather
Shared threadSeveral agents speak in one threadGroup Chat
Inner sessionAn agent triggers an internal sub-conversationNested Chat
Shared stateAgents don't chat directly; they read/write a workspaceBlackboard
Event streamCommunication via topic / queue / event busEvent Bus / Pub-Sub

3. Decision making

TypeHow a decision is reachedRepresentative patterns
Manager picksCoordinator chooses the next stepSupervisor, Router
Reviewer judgesCritic / verifier scores qualityGenerator-Critic
IterativeLoop until a condition is metRefinement Loop
AdversarialPro/con sides argue; a judge concludesDebate / Judge
GroupMultiple agents vote or ensembleVoting / MoA
MarketAgents bid, price, or negotiateMarket / Contract Net

4. Execution environment

TypeEmphasisRepresentative patterns
Role-drivenAgents have professional roles and SOPsRole-playing / Virtual Company
Environment isolationAgents run in distinct workspaces / sandboxesWorkspace Isolation
Human-coupledA human is a special agent — approves, correctsHuman-in-the-loop
Environment-mediatedAgents collaborate by modifying shared environmentStigmergy
Simulated learningAgents interact or train in a simulated environmentSocial Simulation / MARL

5. Protocol interconnect

ProtocolWho connects to whomUse
MCPAgent ↔ Tool / Resource / PromptStandardized exposure of tools, data, prompts
A2AAgent ↔ AgentCross-framework, cross-vendor agent collaboration
ACP / ANPAgent ↔ Agent / NetworkAgent communication, identity, network discovery
Agent Client ProtocolIDE / Client ↔ Coding agentStandardized link between editors and coding agents

In closing

Don't ask "which multi-agent pattern is strongest" — ask:

  • Does the task need decomposition?
  • Can the subtasks run in parallel?
  • Does a specialist need to own the conversation?
  • Is review and rollback required?
  • Is cross-system interop required?
  • Is long-lived state and observability required?

If the answers are mostly "no," don't reach for multi-agent yet. A single agent with the right tools and the right context is often more reliable.