Multi-Agent & Swarms

For ownersSeveral agents splitting one job. Worth it for big jobs, with more parts to check.

Coordination, emergence, and collective intelligence.

25 lessons

Start with lesson 1

Lesson 1 of 25, ~60 minutesWhy Multi-Agent?

The lessons

  1. 01Why Multi-Agent?Multi-agent systems split a big job across several AI agents, each with its own focus and working memory, instead of asking one agent to research, build and review everything at once.Learn · ~60 minutes
  2. 02Heritage of FIPA-ACL and Speech ActsFIPA-ACL was an early standard for how software agents talk, with message types like request, inform and propose. Today's AI agent protocols are rebuilding many of the same ideas with JSON and plain language.Learn · ~60 minutes
  3. 03Communication ProtocolsAgent protocols are shared rules for how AI agents talk to tools and to each other. MCP connects an agent to tools and data, A2A lets agents hand tasks to other agents, and newer ones focus on audit trails and identity.Build · ~120 minutes
  4. 04The Multi-Agent Primitive ModelThis lesson boils multi-agent design down to four building blocks: the agent, the handoff between agents, the shared state they read and write, and the orchestrator that decides who works next. Popular frameworks are different mixes of these four.Learn · ~60 minutes
  5. 05Supervisor / Orchestrator-Worker PatternIn the supervisor pattern, one lead agent plans the work, hands pieces to specialist worker agents running in parallel, then gathers and combines their results into one answer.Learn + Build · ~75 minutes
  6. 06Hierarchical Architecture and Its Failure ModeHierarchical agent systems stack supervisors: a top manager agent directs sub-managers, who in turn direct worker agents, much like an org chart.Learn + Build · ~60 minutes
  7. 07Society of Mind and Multi-Agent DebateMulti-agent debate has several AI copies answer the same question, read each other's answers, point out mistakes and revise over a few rounds until they settle on one answer.Learn + Build · ~60 minutes
  8. 08Role Specialization — Planner, Critic, Executor, VerifierRole specialization gives each agent one job, such as planner, executor, critic or verifier, often with step-by-step procedures written into its instructions, much like a small software team.Learn + Build · ~60 minutes
  9. 09Parallel / Swarm / Networked ArchitecturesIn a swarm or network design there is no central boss. Agents pull jobs from a shared queue or event stream, work on their own schedule and post results back for others to use.Learn + Build · ~75 minutes
  10. 10Group Chat and Speaker SelectionGroup chat orchestration puts several agents in one shared conversation, and a selector, which can be a model, a fixed rotation or custom code, picks who speaks next.Learn + Build · ~60 minutes
  11. 11Handoffs and Routines — Stateless OrchestrationHandoffs and routines are a minimal way to build agent teams: each agent is a set of instructions plus tools, and passing work to another agent is just a tool call that returns that agent.Learn + Build · ~60 minutes
  12. 12A2A — The Agent-to-Agent ProtocolA2A, the Agent2Agent protocol, is an open standard started by Google for agents to find each other and share tasks. Each agent publishes a card describing its skills, and tasks move through tracked states with attached results.Learn + Build · ~75 minutes
  13. 13Shared Memory and Blackboard PatternsShared memory lets agents work from common notes. In a message pool, every agent sees every message. On a blackboard, agents subscribe only to the updates they care about.Learn + Build · ~75 minutes
  14. 14Consensus and Byzantine Fault Tolerance for AgentsByzantine fault tolerance is a way for a group to reach agreement even when some members are wrong or dishonest. This lesson applies it to AI agents, for example by weighting each vote by confidence.Learn + Build · ~75 minutes
  15. 15Voting, Self-Consistency, and Debate TopologyVoting runs several agents or several samples on the same question and takes the majority answer. Mixing different models and prompts keeps everyone from making the same mistake, and how agents are connected changes results.Learn + Build · ~75 minutes
  16. 16Negotiation and BargainingThis lesson covers AI agents that bargain over prices, resources and terms, what research shows about how well they negotiate, and which tactics give them an edge.Learn + Build · ~75 minutes
  17. 17Generative Agents and Emergent SimulationGenerative agents are AI characters in a simulated world that keep a memory log, reflect on it and plan their days. In a well-known study, a small town of them organized a party from one seeded idea.Learn + Build · ~75 minutes
  18. 18Theory of Mind and Emergent CoordinationTheory of mind is the ability to reason about what someone else knows or believes. Research shows AI agents can do some of this when prompted to, which helps them coordinate, but it does not happen automatically.Learn + Build · ~75 minutes
  19. 19Swarm Optimization for LLMs (PSO, ACO)Swarm optimization borrows from bird flocks and ant colonies to search for good solutions. This lesson shows those methods used with language models to tune prompts, combine models and route tasks.Learn + Build · ~75 minutes
  20. 20MARL — MADDPG, QMIX, MAPPOMulti-agent reinforcement learning trains several agents together through trial and error. Methods like MADDPG, QMIX and MAPPO let agents learn with a shared view during training, then act on their own.Learn · ~90 minutes
  21. 21Agent Economies, Token Incentives, ReputationAgent economies are systems where AI agents earn, spend and build reputations, with ways to identify agents, pay them and reward the ones that contributed. Crypto networks and research projects are testing versions of this.Learn · ~75 minutes
  22. 22Production Scaling — Queues, Checkpoints, DurabilityRunning many agent workflows at once needs work queues and saved checkpoints, so any worker can resume any job after a crash without repeating side effects like emails or charges.Learn + Build · ~75 minutes
  23. 23Failure Modes — MAST, Groupthink, Monoculture, Cascading ErrorsThis lesson catalogs how multi-agent systems fail: vague specifications and roles, breakdowns in coordination and missing checks, plus groupthink, where agents built on one model repeat each other's errors.Learn · ~75 minutes
  24. 24Evaluation and Coordination BenchmarksThis lesson reviews benchmarks for testing multi-agent systems: how agents are connected, how they handle split information, and how they perform in specialized fields, enterprise setups and software work.Learn · ~75 minutes
  25. 25Case Studies and the 2026 State of the ArtThis lesson walks through real multi-agent systems end to end: a lead-and-workers research system, role-based software teams, and an open-source local agent with its own agent-only social network.Learn (capstone) · ~90 minutes

Lessons from AI Engineering from Scratch by Rohit Ghumare and contributors, MIT license. Notes for owners by Lozo. License

Next step

Rather have it built for you?

Reading is optional. Tell me what's stuck and I'll build it into the business you already run.

Tell me what's stuck
  • A real conversation, not a pitch
  • You keep your accounts
  • Nothing sends without your yes