What Is Multi-Agent System? A Plain-English Definition
A multi-agent system is an application in which several AI agents, each with its own role, tools and context, work together on one task. Usually a coordinating agent breaks the job down, hands pieces to specialised agents and combines what they return. The term is older than today's language models; in current usage it almost always means a system built from several model calls.
How does a multi-agent system work?
One agent plans the work, launches other agents with their own instructions and toolsets, then collects and merges their results. Each agent runs in a separate context window and only knows what it is explicitly given, so the quality of the hand-offs largely decides the quality of the outcome.
Take a support ticket about a faulty machine. One agent classifies the ticket and extracts the serial number, another searches service manuals, a third checks the warranty status in the ERP system through tool use. The coordinator turns this into a draft reply for a service engineer to review.
When does splitting the work help?
Splitting helps when a task breaks into independent parts that can run at the same time, such as researching many sources. It hurts when steps depend on each other: agents make conflicting assumptions, and merging their output creates new errors.
Anthropic describes a research system with a lead agent and parallel subagents; according to Anthropic, it outperformed a single agent by 90.2% on the company's internal research evaluation. The same report is open about the cost: multi-agent systems used about 15 times as many tokens as chat interactions. It also says that tasks where all agents need the same context, and most coding tasks, are a poor fit today.
Cognition argues the other side: agents should share full traces, not just messages, because every action carries implicit decisions, and parallel agents make decisions that clash. Its recommendation for most cases is a single agent working in a linear thread.
What should you watch out for?
Start with one agent and split only when a measurable limit justifies it, such as a context window that keeps overflowing or work that genuinely runs in parallel. Every extra agent adds cost, new failure points and testing effort.
- Know the failure patterns: a study of more than 1,600 traces from seven frameworks (Cemri et al.) groups 14 failure modes into three categories: system design, misalignment between agents, and task verification.
- Write precise hand-offs: each sub-agent needs a goal, an output format, a list of allowed tools and a clear stopping point.
- Measure, don't guess: compare single- and multi-agent versions with the same evals, including token cost per task.
- Keep people accountable: approvals for quotes, orders or customer messages stay with a person (human-in-the-loop).
For examples of which agent structure suits a given business process, see AI agents for business functions.
Frequently asked questions
Should every serious AI project use multiple agents?
No. A single agent with good tools is easier to test, cheaper to run and often more reliable. Add agents only when you hit a concrete limit such as context size or work that genuinely runs in parallel.
What is the difference between a multi-agent system and a workflow?
In a fixed workflow, code decides the order of steps. In a multi-agent system, at least one model decides which agents to start and what to hand them.
How do I know whether splitting the work helped?
Run the same evaluation set against a single-agent and a multi-agent version and compare quality, latency and cost per task.