# Multi-Agent Wiki

> A working reference for multi-agent interaction patterns, classification, and engineering implementation. Each pattern answers four questions: what problem it solves, what its communication structure is, how to implement it in a real system, and when *not* to use it.

Site sections:

- [Multi-Agent Wiki](/) — A systematic reference for multi-agent interaction patterns, taxonomies, and engineering implementation.
- [Taxonomy](/taxonomy) — A five-dimensional taxonomy of multi-agent patterns
- [Decision Matrix](/decision-matrix) — Pick a multi-agent pattern by task characteristics

## Patterns

- [Supervisor / Manager](/patterns/supervisor-manager) — A primary agent plans, routes, and synthesizes; specialists execute subtasks.
- [Agents-as-tools](/patterns/agents-as-tools) — Wrap specialist agents as tools; the host agent keeps conversation control.
- [Handoff / Router / Transfer](/patterns/handoff-router) — The current agent transfers conversation control to another agent.
- [Hierarchical Decomposition](/patterns/hierarchical-decomposition) — Multi-level manager-worker tiers — upper levels decompose, lower levels execute, recursively if needed.
- [Graph / State Machine / Workflow](/patterns/graph-workflow) — Define the flow with an explicit graph or state machine, not by letting the LLM improvise.
- [Peer-to-peer / Swarm](/patterns/peer-swarm) — No fixed center; agents self-organize through direct messaging, shared environment, or dynamic handoff.
- [Coordinator / Dispatcher](/patterns/coordinator-dispatcher) — Distributes requests across agents, workflows, and tools — owns task state, retry, timeout, and routing.
- [Sequential Pipeline](/patterns/sequential-pipeline) — Tasks pass through agents in a fixed order; each step's output feeds the next.
- [Parallel Fan-out / Gather](/patterns/parallel-fanout-gather) — Split a task across many agents in parallel; an aggregator merges the results.
- [Group Chat / Meeting](/patterns/group-chat) — Multiple agents share one thread; a moderator or selector picks who speaks next.
- [Nested Chat / Inner Team](/patterns/nested-chat) — An agent runs an internal multi-agent sub-conversation before replying outward.
- [Blackboard / Shared Memory / Workspace](/patterns/blackboard-shared-memory) — Agents collaborate indirectly through shared state, knowledge, a task board, or a workspace.
- [Event Bus / Pub-Sub](/patterns/event-bus-pubsub) — Agents communicate asynchronously through events, topics, or queues — not direct calls.
- [Debate / Judge / Voting](/patterns/debate-judge) — Multiple agents argue; a judge or vote concludes.
- [Generator-Critic / Verifier](/patterns/generator-critic) — One agent produces; another critiques, verifies, scores, or proposes edits.
- [Refinement Loop / Evaluator-Optimizer](/patterns/refinement-loop) — Generate → evaluate → revise, until an exit condition or budget is reached.
- [Market / Auction / Contract Net](/patterns/market-auction-contract-net) — Tasks and resources are allocated via bidding, pricing, or contract-net protocol.
- [Mixture-of-Agents / Layered Ensemble](/patterns/mixture-of-agents) — Multiple models or agents stack in layers; each layer reads multiple prior outputs and improves them.
- [Voting / Ensemble](/patterns/voting-ensemble) — Independent candidate answers from multiple agents; a vote, score, or verifier picks the final.
- [Clarification-at-edge / Ask-before-act](/patterns/clarification-at-edge) — Insert a clarification step at agent-agent handoff boundaries or before uncertain actions.
- [Role-playing / SOP / Virtual Company](/patterns/role-playing-sop) — Agents play product, architect, dev, QA roles bound by a documented SOP.
- [Human-in-the-loop](/patterns/human-in-the-loop) — A human is a special agent — for approval, correction, routing, interruption, or final decisions.
- [Workspace / Sandbox Isolation](/patterns/workspace-isolation) — Each agent runs in its own workspace, git worktree, container, or sandbox.
- [Stigmergy / Environment-mediated Collaboration](/patterns/stigmergy-environment-mediated) — Agents collaborate by leaving traces in the environment; others observe and react.
- [Protocol-mediated Agent Network](/patterns/protocol-mediated) — Connect tools, agents, clients, and platforms via standard protocols — MCP, A2A, ACP, Agent Client Protocol.
- [Composite Pattern](/patterns/composite-pattern) — Real production systems combine pipeline, parallel, handoff, critic, HITL, blackboard, and protocol layers.
- [Coalition / Federation / Holonic Organization](/patterns/coalition-federation-holonic) — Agents form temporary coalitions, teams, federations, or holons around tasks.
- [Social Simulation / Agent Society](/patterns/social-simulation) — Simulate a population, organization, or society with long-term memory, relationships, and emergent behavior.
- [MARL / CTDE](/patterns/marl-ctde) — Multi-Agent Reinforcement Learning with centralized training and decentralized execution.

## Implementation

- [Production Multi-Agent Runtime](/implementation/production-runtime) — An architecture for internal coding-agent platforms and agent backends.
- [Orchestrator Implementation Guide](/implementation/orchestrator) — Scheduler, router, and state machine for a multi-agent platform.
- [Observability and Event Model](/implementation/observability) — Trace, event, and metrics design for a multi-agent platform.
- [Safety, Permissions, and Guardrails](/implementation/safety-guardrails) — Security boundary design for multi-agent platforms.
- [Wiki Content Model](/implementation/content-model) — How to author a page for each multi-agent pattern.
- [Pattern Page Template](/implementation/pattern-page-template) — Use this template when adding a new multi-agent pattern page.

## Reference

- [Glossary](/reference/glossary) — Multi-Agent common terms
- [References](/reference/references) — Public sources used by the Multi-Agent Wiki

## Full content

- [/llms-full.txt](/llms-full.txt) — every wiki page concatenated as plain markdown.

## Source

- [GitHub](https://github.com/fuergaosi233/multiagent-explorer)