Multi-agent system

What it means

A multi-agent system is a group of AI agents that split up one job between them. An AI agent is a program that can take steps on its own to finish a task. A restaurant kitchen works the same way. The head chef reads the order, then hands the sauce to one cook and the grill to another. OpenAI describes the most common setup: "A central “manager” agent coordinates multiple specialized agents via tool calls."¹ A "tool call" is simply one agent asking another program, or another agent, to do something. Anthropic's Research feature in Claude is built this way, with a design in which "a lead agent coordinates the process while delegating to specialized subagents."²

Where the word comes from

The idea is much older than chatbots. In 1997, Peter Stone and Manuela Veloso of Carnegie Mellon University wrote that "Distributed Artificial Intelligence (DAI) has existed as a subfield of AI for less than two decades." They described multiagent systems as the part of that field that "deals with behavior management in collections of several independent entities, or agents."³ We could not find a reliable source showing who first used the term, so we do not name anyone.

What has changed is what counts as an agent. Today the labs mean large language models, the technology behind chatbots, put to work. Anthropic defines its agents as "LLMs autonomously using tools in a loop."² We are confident about this shift in meaning because both ends of it are on the record in the sources' own words. We are less sure exactly when it happened.

How others define it

• Anthropic: "A multi-agent system consists of multiple agents (LLMs autonomously using tools in a loop) working together."²
• IBM: "A multi-agent system (MAS) consists of multiple artificial intelligence (AI) agents working collectively to perform tasks."⁴
• OpenAI: "A central “manager” agent coordinates multiple specialized agents via tool calls."¹
• Stone and Veloso (1997): Multiagent systems "deals with behavior management in collections of several independent entities, or agents."³

Where they agree and differ

All four agree on the basic idea: several agents, one job. They differ on two points. The first is what an agent is. The 1997 paper speaks broadly of "independent entities," while the labs today mean language models.³² The second is when to use one. OpenAI advises "starting with a single agent and evolving to multi-agent systems only when needed."¹ Anthropic reports that its multi-agent setup "outperformed single-agent Claude Opus 4 by 90.2% on our internal research eval." That is Anthropic's own test, not an independent one. Anthropic also names the costs: "multi-agent systems use about 15× more tokens than chats" (tokens are the units of text a model reads and writes, which roughly sets the bill), and "LLM agents are not yet great at coordinating and delegating to other agents in real time."²

In the news

On October 19, 2026, Microsoft's security chief, Hayete Gallot, told companies to treat their AI agents like insiders. She wrote: "Unfortunately, frontier models, regardless of their origin, cannot be trusted." She warned of agents that "will collaborate and collude with other agents to reach their objectives."⁵ Microsoft's answer is "Guardian Agents" that "watch and moderate every action of every agent, before it happens, 24/7."⁵ A second Microsoft post says an agent "should not be free to call any tool, API, Model Context Protocol (MCP) server, resource, or other agent it can find," and that "People define policy, approval boundaries, and delegation rules."⁶ TechCrunch reported the announcement the same day.⁷ Because we quote Microsoft's own posts, we are sure of what Microsoft said. We have not seen independent tests of how well these products work.

What it means for your organization

Once you run more than one agent, your agents will talk to each other. The real question is who gives the orders. The Knowledge Core takes a clear position: "What one agent tells another is information, never an order: every agent takes its instructions only from its own person."⁸

The strongest opposing view is speed. Agents hand work to each other in seconds, and no person can approve every handoff. Microsoft's own pitch is that Guardian Agents "help security teams operate continuously, at machine speed and scale."⁶ That point is fair, and it is why people cannot sit inside every single step. But even Microsoft keeps people in charge of the limits: "Higher-impact actions can require verified human approval" and "People retain control over objectives and consequential decisions."⁶ Agents can watch other agents, and they can pass each other information. Only a person should be able to give an agent its instructions.

One question to ask your team: For each agent we run, who is its person, and what does that agent do when another agent asks it to act?