Multiagent Systems

Multiagent systems is the branch of artificial intelligence that studies what happens when many autonomous programs, robots, or simulated actors share an environment and must get along.

15 frameworks, from 1971 to today · Artificial Intelligence · Computer Science

still a working view ended · each leads to its entry below

Overview

Multiagent systems is the branch of artificial intelligence that studies what happens when many autonomous programs, robots, or simulated actors share an environment and must get along. An agent here means a system that perceives its world and acts on it to pursue its own goals; a multiagent system is any collection of such agents whose actions affect one another12. What is at stake is coordination without a single controller: how independent decision-makers can divide work, share information, avoid collisions and deadlock, negotiate over scarce resources, and cope with others who may be selfish, adversarial, or simply different1. The field's history is a long convergence of three streams — distributed problem-solving, the economics of strategic interaction, and the engineering of robot and software societies — that only became one recognizable discipline in the 1990s.

Distributed problem solving and the birth of DAI

The earliest strand grew out of 1970s distributed computing and speech-understanding research. The Hearsay-II speech system (Carnegie Mellon, late 1970s) split recognition among many "knowledge sources" that posted and read partial hypotheses on a shared blackboard — an early demonstration that loosely coupled specialists could solve a problem no single module could. Around the same time, work on distributed artificial intelligence (DAI) asked how a network of problem-solvers could cooperate on a task when no one node had the whole picture3. The Contract Net Protocol, proposed by Reid Smith in 1980, was the landmark: a manager announces a task, contractors bid on it, and the manager awards the contract4. It gave a concrete, reusable mechanism for task allocation by negotiation, and it framed the central DAI question — how to decompose work and share results — in terms practitioners could implement43.

Reactive robots and the behavior-based turn

A second stream came from robotics. Classical planning assumed an agent built a world model, planned, then acted. Rodney Brooks argued in the mid-1980s, most famously in "A Robust Layered Control System for a Mobile Robot" (1986), that intelligence could emerge from layered reactive behaviors — simple rules responding directly to sensor input — with no central model at all. His subsumption architecture let higher layers override lower ones, and robots built this way (the Genghis and Herbert machines of the late 1980s) walked and collected objects with remarkable robustness. The lesson for multiagent work was double-edged: simple local rules can produce coherent global behavior, but purely reactive agents struggle with anything requiring memory or long-range planning. That tension pushed the field toward hybrid designs and, later, toward learning.

Swarms, simulation, and agent-based modeling

The idea that local rules generate global order had a parallel in the study of collective behavior5. Craig Reynolds's "boids" (1987) showed that three simple steering rules — separation, alignment, cohesion — reproduced realistic flocking. In 1989 Gerardo Beni and Jing Wang coined swarm intelligence for the emergent collective behavior of decentralized robot groups, and the 1990s brought influential optimization methods modeled on insect colonies, notably Marco Dorigo's ant colony optimization (early 1990s) and Russell Eberhart and James Kennedy's particle swarm optimization (1995). Meanwhile, in the social sciences, agent-based modeling — associated with researchers such as Joshua Epstein and Robert Axtell, whose Sugarscape model appeared in the mid-1990s — used populations of simple simulated agents to study how macro patterns like segregation or wealth inequality arise from micro decisions. Thomas Schelling's earlier segregation model (1971) is the classic ancestor. These communities shared the swarm intuition but pursued different goals: optimization, control, and social explanation.

Game theory and mechanism design enter AI

Strategic interaction gave the field its mathematical backbone6. Game theory, developed by von Neumann and Morgenstern in the 1940s and extended by John Nash's equilibrium concept (1950), describes how rational agents choose when each one's payoff depends on the others67. Its central solution idea, the Nash equilibrium, is a profile of strategies from which no single agent can gain by unilaterally changing course7. From the late 1980s, AI researchers began importing these tools to analyze and design agent interactions. Mechanism design — "reverse game theory," which asks what rules produce good outcomes given that agents act in their own interest — became central to auctions and resource allocation8. The 1994 paper by Jeffrey Rosenschein and Gilad Zlotkin, Rules of Encounter, applied negotiation and game theory directly to automated agents, and the 1995 appearance of the Journal of Artificial Intelligence Research paper on the subject by Tuomas Sandholm and Victor Lesser on coalition formation signaled a maturing research program. Auction-based allocation later became a standard tool for distributed scheduling and, in the 2000s, for allocating computing resources.

BDI agents and the software-agent boom

The 1990s also saw agents move from robots and simulations into software. The belief–desire–intention (BDI) model, developed by Michael Bratman in philosophy and adapted to computing by Anand Rao and Michael Georgeff in the early 1990s, described an agent by its beliefs about the world, its desires (goals), and the intentions it has committed to. It gave programmers a principled architecture for practical reasoning, and it anchored a family of agent programming languages and platforms91011. The decade's enthusiasm for "software agents" — interface agents, information agents, mobile agents — produced standards for how agents should talk. KQML (Knowledge Query and Manipulation Language), developed in the early 1990s under Tim Finin and colleagues, and its successor FIPA-ACL, standardized by the Foundation for Intelligent Physical Agents from the late 1990s, defined message types such as inform, request, and propose, with a semantics grounded in speech-act theory. These agent communication languages let independently built agents interoperate, though the gap between formal semantics and working implementations remained a persistent complaint.

Cooperation, constraints, and teamwork

Running alongside was a body of work on how agents coordinate when they must act together. Distributed constraint reasoning — formalized in the 1990s by researchers including Makoto Yokoo and, in a widely cited 2000 survey, by Pragnesh Jay Modi and colleagues — treated coordination as a constraint-satisfaction problem spread across agents, each holding part of the problem and communicating to find a globally consistent assignment. Teamwork research, notably the STEAM system by Milind Tambe (late 1990s), addressed how agents in a team maintain shared plans and adapt when members fail. These threads established that coordination could be studied as an algorithmic problem with measurable communication costs, not just a design philosophy.

Multiagent reinforcement learning

A third major stream asked how agents learn to coordinate through experience12. Reinforcement learning — learning by trial, error, and reward — had been developed for single agents, but the multiagent case is harder because each agent's environment includes the others, who are also learning. Work in the 1990s and 2000s, associated with researchers such as Michael Littman, Caroline Claus, and Craig Boutilier, formalized multiagent RL and showed that naive independent learners can oscillate or converge to poor outcomes. The field developed techniques for joint-action learning, communication, and credit assignment, and the 2010s brought deep multiagent RL, in which neural networks approximate policies for many interacting agents. A parallel line, opponent modeling, studies how an agent can infer the strategy of others and exploit it — from simple pattern recognition in repeated games to equilibrium-based reasoning. The tension between learning a robust general policy and exploiting a specific opponent's weaknesses remains a live design choice.

The present: learning, language models, and open questions

Today the field's center of gravity has shifted toward learning-based coordination. Deep multiagent reinforcement learning underpins work on robot teams, traffic control, and game-playing, while game-theoretic mechanism design remains the toolkit for auctions and marketplaces. The most recent turn is the use of large language models as agents: since the early 2020s, systems have been built in which several LLM-driven agents converse, assign roles, and critique one another to solve tasks — a return, in new form, to the old dream of agents that coordinate through language. Whether such systems genuinely coordinate or merely produce the appearance of it is an open question, as are the field's long-standing ones: how to guarantee that self-interested agents reach socially good outcomes, how to scale coordination to thousands of heterogeneous agents, and how to verify that a multiagent system behaves safely when no one designed its global behavior. Those questions descend directly from the blackboard systems, contract nets, and boids of the field's first decades.

References

  1. Multi-agent system (introduction, Characteristics, Research), Wikipedia
  2. Artificial intelligence, Wikipedia
  3. Distributed artificial intelligence (Development), Wikipedia
  4. Contract Net Protocol, Wikipedia
  5. Swarm intelligence, Wikipedia
  6. Game theory (introduction, History), Wikipedia
  7. Nash equilibrium (introduction, History), Wikipedia
  8. Mechanism design, Wikipedia
  9. Belief–desire–intention software model, Wikipedia
  10. 3APL, Wikipedia
  11. AgentSpeak, Wikipedia
  12. Multi-agent reinforcement learning, Wikipedia

The frameworks of Multiagent Systems

Each framework is a way of working the field has used: when it held, and the question it answers.

  1. 1971-Present

    Agent-based model

    How do simple agents produce social patterns?

    Fly to its tower in Nousfarer

  2. 1975-1995

    Distributed Artificial Intelligence

    How can distributed problem-solvers cooperate without a controller?

    Fly to its tower in Nousfarer

  3. 1980-Present
  4. 1986-2000

    Behavior-based robotics

    Can intelligence emerge from layered reactive behaviors?

    Fly to its tower in Nousfarer

  5. 1987-Present
  6. 1988-Present

    Game-theoretic multiagent analysis

    How do self-interested agents choose when payoffs interact?

    Fly to its tower in Nousfarer

  7. 1989-Present

    Swarm intelligence

    How can colonies of simple agents solve hard problems?

    Fly to its tower in Nousfarer

  8. 1990-Present

    Mechanism design

    What rules make self-interested agents produce good outcomes?

    Fly to its tower in Nousfarer

  9. 1991-2010

    Belief–desire–intention software model

    How should an agent reason about beliefs, goals and commitments?

    Fly to its tower in Nousfarer

  10. 1993-2010

    Agent Communications Language

    How should independent agents exchange messages?

    Fly to its tower in Nousfarer

  11. 1993-Present

    Distributed constraint reasoning

    How can agents satisfy constraints held across them?

    Fly to its tower in Nousfarer

  12. 1995-Present

    Multi-agent reinforcement learning

    How do agents learn to coordinate through experience?

    Fly to its tower in Nousfarer

  13. 1995-Present

    Opponent modeling

    How can an agent infer and exploit another's strategy?

    Fly to its tower in Nousfarer

  14. 1996-2010

    Teamwork and joint intention

    How do teams maintain shared plans when members fail?

    Fly to its tower in Nousfarer

  15. 2022-Present

    LLM-based multiagent systems

    Can language-model agents genuinely coordinate through talk?

    Fly to its tower in Nousfarer

How the ideas connect

Each arrow runs from an earlier idea to the later one it shaped. The list below says what passed between them.

Agent-based modelDistributed Artificial IntelligenceCooperative Teamwork and Distributed …Behavior-based roboticsSwarm Intelligence and Agent-Based Mo…Game-theoretic multiagent analysisSwarm intelligenceMechanism designBelief–desire–intention software modelAgent Communications LanguageDistributed constraint reasoningMulti-agent reinforcement learningOpponent modelingTeamwork and joint intentionLLM-based multiagent systems

Beyond the field

Where Multiagent Systems's ideas came from in other fields, and what they led to there: the strongest 12 of 58 links.

A city in Nousfarer: a red-carpeted stair climbing terraces to a crowned tower on an acropolis, more towers round it, forest and snowy mountains beyond.
A city in Nousfarer, the game played in the browser.

Multiagent Systems is a planet in Nousfarer

Its city has a tower for each of these 15 frameworks. Land, walk in, and climb one concept at a time.

Land on Multiagent Systems

Related fields

Fields its ideas are linked to

More in Artificial Intelligence

See where Multiagent Systems sits on the Map of Human Knowledge