Multi-agent system

What it is

A multi-agent system is a computerised system composed of multiple interacting intelligent agents^[1]. The point of putting several agents together rather than one is capability: such systems can solve problems that are difficult or impossible for a single agent or a monolithic system^[1]. The field is related to agent-based modelling, but it is oriented towards engineering applications rather than towards simulating phenomena for study^[1].

Each agent is typically a self-contained piece of software that perceives something of its environment and acts on it; the interesting part is what happens between them.

How the agents work together

Because no single agent sees the whole problem, coordination mechanisms are central. Standard approaches include the contract net protocol, in which tasks are announced and taken on by agents that can do them; market mechanisms, which allocate work through something like bidding and prices; and stigmergy, a bio-inspired form of indirect coordination in which agents leave traces in a shared environment that others respond to^[1].

Agents also need a shared language. Agent communication languages include KQML, ACL and the FIPA standards^[1]. Beyond explicit protocols, systems can rely on self-organisation, which can produce emergent behaviour — patterns of activity that no individual agent was programmed to produce^[1].

Where the field came from

The idea has a long literature. Hewitt and Inman worked on distributed artificial intelligence in 1991, and Singh published a treatment titled "Multiagent Systems" in 1994^[1]. Consolidating textbooks followed: Weiss edited a foundational book in 1999, Wooldridge published "An Introduction to MultiAgent Systems" in 2002, and Shoham and Leyton-Brown published on multiagent systems in 2009^[1]. These works turned a scattered set of ideas into something that could be taught and built on.

Large language models as agents

The most recent wave puts large language models inside each agent. In 2023, Li and colleagues introduced CAMEL, a large-language-model multi-agent framework^[1]. Researchers have reported that language-model agents can develop communication spontaneously, and that this brings its own problems: semantic drift, and a gap between what can be observed and what can be interpreted, as reported in Science in 2026^[1].

Two frameworks from 2023 show what this looks like in practice. AutoGen, first submitted in August 2023, is an open-source framework for building LLM applications out of multiple agents that converse with each other to accomplish tasks^[2]. Its agents are customizable and conversable, and can operate in modes that combine LLMs, human input and tools^[2]. Both natural language and computer code can be used to program conversation patterns for different applications^[2].

Generative Agents, first submitted in April 2023, describes an architecture that extends a large language model to store a complete record of an agent's experiences in natural language, synthesize those memories into higher-level reflections, and retrieve them dynamically to plan behaviour^[3]. The authors instantiated 25 generative agents in an interactive sandbox environment inspired by The Sims^[3].

Where they are used

Applications include online trading, disaster response, target monitoring, games, transport, logistics, power grids and autonomous driving simulation such as Waymo Carcraft^[1]. The common thread is a setting with many parties, partial information and a need to act under change.

Limitations

The approach has real costs. Reported limitations include weak scalability, non-determinism arising from protocol drift, a lack of formal verification, and the difficulty of overseeing a system whose behaviour comes out of many interacting parts^[1]. The same openness that makes multi-agent systems capable is what makes them hard to guarantee.