A multi-agent system is several AI agents working on one job. Usually a lead agent, the orchestrator, splits the work, hands pieces to helper agents called workers or subagents, and combines what they send back. It pays off when the work splits into independent parts, such as checking twenty sources at once. It backfires when every step depends on the last.
Recent research gives a simple rule: split work that can run in parallel, and keep work that runs in sequence with one agent. This guide covers when multi-agent systems help, when they hurt, what documented systems actually do, and how to design a simple one.
- A multi-agent system puts several agents on one job, most often an orchestrator that delegates to workers and merges their results.
- It helps with broad work that splits cleanly: research, searching a large codebase, reviewing many files.
- It hurts on tightly linked work, where agents that cannot see each other’s choices make conflicting ones.
- It costs more. Anthropic measured multi-agent runs using about 15 times the tokens of a chat.
- Start with one agent. Add a worker only when a side task floods its context or clearly runs in parallel.
What is a multi-agent system?
If you are new to agents, start with what one is: a model using tools in a loop toward a goal. A multi-agent system runs several of those loops and connects them.
That is the most common design, but not the only one. Here are the four patterns you will meet:
| Pattern | How it works | Good for |
|---|---|---|
| Orchestrator and workers | A lead agent decides the subtasks at run time and delegates them | Research, broad searches, reviews |
| Handoff | A triage agent passes the whole conversation to a specialist | Support desks, routing requests |
| Pipeline | Each agent’s output becomes the next one’s input, in a fixed order | Draft, edit, format |
| Checker | One agent produces, another critiques, and they loop | Code review, fact-checking |
OpenAI’s Agents SDK names the first two “agents as tools” and “handoffs”. The difference is who stays in charge. As a tool, a specialist answers the manager and the manager keeps talking to the user. After a handoff, the specialist takes over.
When several agents beat one
Extra agents help for three reasons.
Breadth. Workers run at the same time, so a question that needs twenty searches finishes faster and covers more ground.
Clean context. Each worker starts with a fresh context window and returns a short summary, not every page it read. The lead agent keeps room to think. Context engineering explains why a crowded window makes models worse.
Focus and limits. A worker can get a narrow brief, a cheaper model or read-only access. A reviewer that cannot edit files cannot break them.
On the right tasks, the measured gains are large, and so is the bill:
When multiple agents make things worse
Google Research tested 180 agent setups across four benchmarks. On PlanCraft, a planning task where each step depends on the last, every multi-agent variant did worse than one agent, by 39 to 70%. Splitting a chain of decisions only adds handoffs where things get lost.
Cognition, the company behind Devin, made the same point in 2025 with two rules: share full context, and remember that “actions carry implicit decisions”. Its example: asked for a Flappy Bird clone, one parallel worker drew a Super Mario style background and another drew a bird that did not match. Neither was wrong alone. Together, the result was unusable.
Errors also spread. In Google’s tests, agents working independently amplified errors 17.2 times, while a central coordinator held that to 4.4 times. A 2025 study of seven multi-agent frameworks sorted their failures into three groups: poor system design, agents misaligned with each other, and weak checking of the final result.
- The work splits into parts that do not depend on each other
- Side tasks would flood the main context with search results or logs
- Workers only read and report, rather than change things
- The answer is valuable enough to justify a much bigger bill
- Each step depends on the one before it
- Several agents would edit the same files or make design choices
- One agent already does the job reliably
- Speed or cost matters more than breadth
Real multi-agent systems, documented
These are systems whose builders published how they work.
A lead agent plans the research and starts parallel subagents, each searching one angle, then a separate agent adds citations. Early versions spawned 50 subagents for simple questions, so Anthropic wrote scaling rules into the prompts: one agent for a simple fact, two to four for a comparison, more than ten for complex research.
Each subagent is a Markdown file with its own instructions, tool list and model. It works in a fresh context and returns only a summary. Built-in ones include Explore, a fast read-only searcher. The docs advise keeping work that shares a lot of context, like planning then building, in the main conversation.
Supports both manager-style agents as tools and handoffs. It also suggests orchestrating in code when you want predictable speed and cost, for example running agents in parallel and chaining their outputs.
Argues for one agent with continuous context, plus compression of old history for long tasks. It notes that Claude Code limited its helpers to answering questions, not writing code, to avoid conflicting work.
The pattern across all four is the same. Helpers that read and report work well. Helpers that make decisions in parallel need a lot of care.
How to design a simple multi-agent system
Start with one agent
Run the job with a single agent first and watch where it struggles. The usual signs are a context full of search results, or independent subtasks done one by one.
Split only the independent parts
Good splits are one worker per source, per folder or per question. Bad splits are frontend and backend of the same feature, or two halves of one document.
Write each worker a full brief
Workers do not see your conversation. Give each one the goal, what to return and in what format, which tools to use, and where its job ends.
Keep decisions in one place
Workers gather and report. The orchestrator decides. If two workers could make conflicting choices, one agent should make them in sequence instead.
Cap the fan-out
Set a maximum number of workers and tool calls, and scale them to the question. A simple lookup needs one agent, not ten.
Check the merged result
Add a final pass that checks the combined answer against the original goal and the sources, before anyone relies on it.
A good worker brief reads like a ticket for a contractor who has never seen your project:
You are one of several research workers. Your only task: find what [company] charges for [product] as of this month. Use the official pricing page first, then the docs. Skip reviews and comparison sites. Return at most 150 words: the prices, the plan names, the page links, and anything you could not confirm. Do not research competitors or features. Other workers cover those. Stop after 10 searches, even if you are not finished, and say what is missing.
How much does a multi-agent system cost?
More, and by design. Anthropic found that agents use about 4 times the tokens of a chat, and multi-agent systems about 15 times. On BrowseComp, a test of finding hard-to-locate facts on the web, token usage alone explained 80% of the differences in performance. Multi-agent systems win partly by spending more.
That makes the math simple. Use them where a better answer is worth several times the cost, such as research that feeds a real decision. Give workers a smaller, cheaper model where the task allows, and cap their number. Our guide to cutting AI costs covers model routing and caching in detail.
FAQ
What is the difference between an AI agent and a multi-agent system?
An agent is one model working in a loop with tools. A multi-agent system connects several agents, each with its own instructions and context, usually under a lead agent that splits the work and combines the results.
Are subagents the same as a multi-agent system?
Yes, subagents are the most common form. A main agent starts helpers for side tasks, and they report back with summaries. The helpers rarely talk to each other.
Do multi-agent systems work for coding?
Partly. Helpers that search a codebase, run tests or review code work well. Several agents writing the same feature in parallel tend to clash, and Anthropic notes that most coding tasks have fewer truly parallel parts than research.
How do agents from different companies work together?
Through a shared protocol. The Agent2Agent protocol, or A2A, is an open standard for one agent to find another, hand it a task and get results back.
How many agents should I use?
As few as the job allows. Start with one, add a worker for each clearly independent part, and set a hard cap.
Next, see ten agentic workflows you can run today, or learn how A2A lets agents talk to each other.
- How we built our multi-agent research system, Anthropic, June 2025
- Towards a science of scaling agent systems: when and why agent systems work, Google Research, January 2026
- Don’t build multi-agents, Cognition, June 2025
- Why do multi-agent LLM systems fail?, Cemri and others, arXiv, October 2025
- Create custom subagents, Claude Code docs
- Orchestrating multiple agents, OpenAI Agents SDK docs
- Building effective agents, Anthropic, December 2024
