Patterns Overview

The patterns below are not mutually exclusive — they compose. A production-grade coding agent might use Supervisor + Parallel Fan-out + Blackboard + Generator-Critic + Human-in-the-loop + MCP, all at once.

#	Pattern	Category	One-liner
1	Supervisor / Manager	Control	A primary agent plans, routes, and synthesizes; specialists execute subtasks.
2	Agents-as-tools	Control	Specialists are exposed as callable tools; the host keeps conversation control.
3	Handoff / Router	Control	The current agent transfers control to another, which takes over the conversation.
4	Sequential Pipeline	Information	Tasks pass through agents in a fixed order; each step's output is the next step's input.
5	Parallel Fan-out / Gather	Information	Split a task across many agents in parallel; an aggregator merges results.
6	Hierarchical Decomposition	Control	Multi-level manager-worker tiers; upper tiers decompose, lower tiers execute, recursively if needed.
7	Graph / State Machine / Workflow	Control	Define the flow with an explicit graph or state machine, not by letting the LLM improvise.
8	Group Chat / Meeting	Information	Multiple agents share one thread; a moderator or selector picks the next speaker.
9	Nested Chat / Inner Team	Information	An agent runs an internal multi-agent sub-conversation before replying outward.
10	Debate / Judge	Decision	Multiple agents take opposing positions, then a judge or vote picks the result.
11	Generator-Critic	Decision	One agent produces; another critiques, verifies, scores, or proposes edits.
12	Refinement Loop	Decision	Generate → evaluate → revise, until an exit condition or budget is met.
13	Role-playing / SOP	Environment	Agents play PM, architect, dev, QA roles bound by a documented SOP.
14	Blackboard / Shared Memory	Information	Agents collaborate indirectly via shared state, knowledge base, task board, or workspace.
15	Event Bus / Pub-Sub	Information	Agents communicate asynchronously through events, topics, or queues — not direct calls.
16	Market / Auction / Contract Net	Decision	Tasks and resources are allocated via bidding, pricing, or contract-net protocol.
17	Peer-to-peer / Swarm	Control	No fixed center; agents self-organize via direct messaging, shared environment, or dynamic handoff.
18	Mixture-of-Agents	Decision	Layered ensemble — each layer reads multiple outputs from the previous and improves them.
19	Human-in-the-loop	Environment	A human is a special agent for approval, correction, routing, interruption, or final decision.
20	Protocol-mediated Network	Protocol	Connect tools, agents, clients, and platforms via MCP, A2A, ACP, Agent Client Protocol.
21	Clarification-at-edge	Decision	Insert a clarification step at agent-agent handoff boundaries or before uncertain actions.
22	Coordinator / Dispatcher	Control	Dispatcher distributes requests to agents, workflows, or tools and owns task state, retry, timeout, routing.
23	Voting / Ensemble	Decision	Independent candidates from multiple agents; a vote, score, or verifier picks the final answer.
24	Composite Pattern	Composition	Real systems combine pipeline + parallel + handoff + critic + HITL + blackboard + protocol layers.
25	Workspace / Sandbox Isolation	Environment	Each agent runs in its own workspace, git worktree, container, or sandbox to avoid concurrent corruption.
26	Stigmergy / Environment-mediated	Environment	Agents leave traces in the environment (issues, todos, diffs, test results); others react to them.
27	Coalition / Federation / Holonic	Organization	Agents form temporary coalitions, federations, or holons — governance, membership, autonomy boundaries.
28	Social Simulation	Simulation	Simulate a population, organization, or society with long-term memory, relationships, emergent behavior.
29	MARL / CTDE	Learning	Multi-Agent Reinforcement Learning; centralized training, decentralized execution.