What is a multi-agent system?
Last reviewed 7 October 2026
A multi-agent system is a set of AI agents that work together on a task, each with its own role, instructions and tools, coordinated by a supervisor agent or by rules for passing work between them. IBM defines it as multiple AI agents working collectively to perform tasks on behalf of a user or another system. The idea predates language models.
Also called: multi-agent systems, MAS, multiagent system, agent team.
Why Multi-agent system matters
Splitting work across agents can make each one simpler. A research agent, a drafting agent and a checking agent each get a short prompt and a few tools, instead of one agent juggling everything. It also lets you give agents different permissions: the agent that reads customer records does not need the tool that sends email.
The cost is real, though. Anthropic reported that in its research system, agents typically used about four times more tokens than chat interactions, and multi-agent systems about fifteen times more. Every handoff is another place for context to get lost. A multi-agent design should earn its place against a single agent with good tools.
How it works
There are two broad shapes. In a centralised design, a supervisor or orchestrator agent breaks the task down, hands parts to worker agents and merges what they return. Anthropic’s research system works this way: a lead agent plans and spawns subagents that explore in parallel. In a decentralised design, agents share information with their neighbours directly, which IBM notes removes the single point of failure but makes coordinated behaviour harder to achieve.
Coordination needs shared state. Agents either read and write a common record of the task, or pass structured messages that say what was done and what is needed next. AutoGen, a framework from Microsoft researchers, models this as a conversation between agents that can combine LLMs, human inputs and tools.
Multi-agent setups suit some tasks and not others. Anthropic found them strong for broad research with heavy parallel work and information that exceeds one context window. It found them weaker where every agent needs the same context or where agents depend heavily on each other, and said most coding tasks have fewer truly parallel parts than research.
Example: a vendor risk review
A risk team asks for a review of a new software vendor. A supervisor agent splits the job three ways: one agent reads the vendor’s security questionnaire, one checks public sources for reported incidents, and one reads the draft contract for data processing terms. Each works with only the tools its job needs.
The three report back with findings and citations. The supervisor merges them, notices that the questionnaire says data stays in one region while the contract allows transfers elsewhere, and flags the conflict at the top of the summary. A risk analyst reviews the summary and decides; no agent approves the vendor.
A single agent could do this in sequence. The multi-agent version is worth it here because the three checks are independent and can run at the same time, and because the contract agent should not have web access.
How Swfte relates to it
Built in the product
Swfte Studio’s Pro tier lists multi-agent orchestration. Agents and fixed workflow steps sit in one visual model, each agent keeps its own identity, permissions and limits, and each handoff appears as a step in the trace. Different agents in one workflow can use different models through Swfte Connect.
The platform does not enforce separation of duties today. You get it by placing approval steps for people, not agents, and by choosing who is assigned to each one.
Related terms
- AI agent orchestration
AI agent orchestration is the coordination of several AI agents, and the tools and people around them, so that they work toward one outcome in a controlled order.
- AI agent
An AI agent is a software program that uses an AI model to decide what to do next and then acts through tools to reach a goal it was given.
- Agentic workflow
An agentic workflow is a business process in which one or more steps are carried out by an AI agent that decides how to complete them, alongside fixed steps and human checkpoints.
- LLM agent
An LLM agent is an agent whose decisions are made by a large language model, which reads a goal, reasons about it and chooses actions such as tool calls in a repeated loop.
Common questions
- When should I use a multi-agent system instead of one agent?
- When parts of the task are independent and can run in parallel, need different permissions, or would overload one agent with too many tools. If steps depend tightly on each other or all need the same context, one agent with good tools is usually cheaper and easier to debug.
- What is a supervisor agent?
- A supervisor, sometimes called an orchestrator or lead agent, receives the overall task, decides how to split it, assigns parts to worker agents and combines their results. It is the most common way to coordinate several agents. It is also a single point of failure, so its instructions and limits need the most care.
- Why are multi-agent systems hard to debug?
- Agents make different decisions on different runs, even with the same input, and a problem in one agent often shows up as odd output in another. You need a trace that records every handoff, what was passed and what came back. Anthropic noted that agents can take different valid paths, which makes fixed expected outputs a poor test.
- Do multi-agent systems cost more?
- Usually yes. Each agent carries its own prompt, tool definitions and context, and the supervisor reads what the workers return. Anthropic reported multi-agent runs using far more tokens than chat in its research system. Set a step and spend limit per run, and use smaller models for workers whose tasks are simple.
Sources
Definitions on this page were read on the sources below on 7 October 2026. Where sources define the term differently, the page says so. The full glossary lists more terms.