A multi-agent system is a system in which many autonomous agents interact with each other to solve problems that are beyond the ability of any single one of them. Each agent is a computational entity - a software program or a robot - that perceives its environment, makes its own decisions in pursuit of its goals, and acts without being centrally controlled at every step. The interactions among agents can be cooperative, like the coordinated foraging of an ant colony, or competitive, like buyers and sellers in a market.
The field predates large language models by decades. Michael Wooldridge’s textbook “An Introduction to MultiAgent Systems,” whose second edition was published by John Wiley and Sons in 2009, became the standard reference and frames the subject as a paradigm for understanding and building distributed systems out of autonomous components. Classic concerns of the field include how agents communicate, how they negotiate and reach agreement, how they coordinate without a central controller, and how to design protocols and incentives so that self-interested agents still produce useful collective behavior. These questions were studied long before a language model could play the part of an agent.
What changed recently is the substrate. When the “agent” became an LLM that can read instructions, reason in natural language, and call tools, the old multi-agent questions reappeared in a new form. Frameworks that assign roles to several LLM instances and let them talk - AutoGen, CAMEL, MetaGPT, ChatDev - are multi-agent systems whose agents happen to be language models, and protocols like A2A revive the communication-and-coordination problem for this generation.
Why business readers should care: when a vendor pitches a “team of AI agents,” the underlying engineering problems - who decides, how they hand off work, what happens when two agents disagree, how to keep costs bounded - are the decades-old hard problems of multi-agent systems, not a solved feature. The maturity of a product is better judged by how it answers those coordination questions than by the number of agents it advertises.