MMOOZZGG.com

Glossary

What is a multi-agent system?

In short

A multi-agent system is a group of AI agents that share one task between them: each has its own role, its own tools and its own part of the work.

AI agents · multi-agent system

MMOOZZGG.com editorsChecked: 1 min read

For a child

Think of how a class makes a wall newspaper. One student looks for interesting facts, another draws, a third invents headlines, and the class leader puts it all on one sheet and checks that nothing is lost. A multi-agent system works the same way: several smart programs work as a team, each has its own job, and one of them watches the overall result.

In plain words

It is hard for one AI agent to hold a big task in full: it gets lost in details and loses the thread. So the work is split. The lead agent breaks the task into parts and hands them to helpers. Each helper gets a short assignment and its own tools, then returns the result. The lead agent combines the answers, asks for a weak part to be redone, and delivers the result. For example, for a market review, one agent collects news, a second collects competitor prices, a third collects customer reviews, and the lead agent writes the overall conclusion.

For business

A team of agents suits processes with several different roles: take a request, check the client, calculate an offer, open a deal. Each role is easier to set up, test and replace on its own than one all-purpose worker. The downside is cost: according to Anthropic, its multi-agent research system uses about 15 times more tokens than an ordinary chat. Complexity grows too: agents pass errors to each other and duplicate work. Before launch, teams calculate the cost of one task, set the rights of each role and keep a log of handoffs. This is how an AI department works: several agents for one department of a company, launched one process at a time.

For a specialist

A multi-agent system (MAS) built on language models consists of several agents that exchange messages and solve a task together. Common topologies are orchestrator-workers, where a lead agent plans and hands out subtasks; a pipeline, where one agent’s result is the next agent’s input; and debate and peer review, where agents challenge each other’s decisions. The AutoGen framework (Wu et al., 2023) described such systems as a programmable conversation of agents that can include people and tools. In June 2025 Anthropic described a research system in which a lead agent launches parallel subagents; on an internal task set it beat a single agent by 90.2%. The price is higher token use, hard debugging and coordination of the shared context. Separate protocols, such as A2A, are being developed for exchange between agents from different vendors.

Example

A real estate agency runs three agents: the first answers requests from the website, the second picks properties from the database, the third books viewings in the agent’s calendar. A coordinator makes sure the client never gets two different answers.

Related terms

Questions and answers

When do you need a multi-agent system, and when is one agent enough? One agent is enough if the task is short and fits one role. Several agents make sense when there are many parts, the parts need different tools, or it pays to run them in parallel.
Why do multi-agent systems cost more? Each agent reads its own instructions and the interim results, so the total amount of processed text grows many times over. Time also goes into setting up handoffs between roles and into finding errors.

Sources

Where companies use it

Not sure whether your company needs this?

AI solutions map
Contact us