Multi-Agent Systems for Beginners: The Playbook on Building AI Teams

Meera works at a small company in Pune. Every Monday her manager asks for a report on three competitors: their prices, their new features and what customers say about them. Meera built an AI helper to write it. One AI agent does everything, one step after another, and it takes forty minutes. Worse, by the end it forgets what it read at the start.

Then a teammate says, “Why not give the job to a small team instead of one worker?” That idea is called multi-agent systems, and this playbook is multi-agent systems for beginners. You will see what they are, when they help, and how to build a simple one safely.

If you have never built an agent, read our AI agents playbook first. It takes five minutes and makes this chapter easier.

Last verified: 5 November 2026. AI tools change fast; check the linked sources for anything newer.

Who this is for: developers and curious beginners who already know what a single AI agent is. Time needed: about 30-45 minutes to read and try the first play. No special framework is required.

What are multi-agent systems, in simple words

Think of a kitchen. One cook can make a whole meal, but a busy restaurant has a head chef who reads the order and several cooks who each handle one dish. The head chef puts the plates together at the end.

A multi-agent system works the same way. It is a group of AI agents where each agent has one clear job, and they pass their results to each other. An agent is an AI model that can use tools and take steps on its own.

The most common design has a lead agent (the head chef) and several worker agents (the cooks). The lead splits the task, the workers do the pieces at the same time, and the lead joins the answers.

What’s new: why teams of agents are talked about now

Anthropic described how it built its own research feature this way: a lead agent plans, then starts several worker agents that search different parts of a question at the same time. In its internal tests this team beat a single agent by 90.2%. It also warned that multi-agent systems use about 15 times the tokens of a normal chat. Tokens are the small pieces of text a model reads and writes, and you pay for them. (Source linked at the end; the test was Anthropic’s own.)

The lesson for beginners is balanced: teams can do more, but they cost more. That is why the first play is about deciding whether you need one.

The playbook at a glance

1. Play 1: Check that one agent is not enough.
2. Play 2: Give each agent one job.
3. Play 3: Pick a simple team shape.
4. Play 4: Write clear hand-off notes.
5. Play 5: Set limits and watch the cost.
6. Play 6: Test the team like a new hire.

Play 1: Check that one agent is not enough

Goal: avoid building a team when one worker is fine. Teams are slower to debug and more costly.

1. Write the task in one sentence.
2. Run it with one agent and note where it fails: too slow, forgets things, or mixes up topics.
3. Ask whether the task splits into parts that do not need each other’s results.
4. If the parts depend on each other all the time, stay with one agent.

Meera’s report has three competitors. Researching each one does not depend on the others, so the job splits naturally. Slow and forgetful are her two real problems.

You are done when: you can name the exact problem a team would fix.

Play 2: Give each agent one job

Goal: make every agent small and focused. A worker with a narrow job makes fewer mistakes, like a cook who only makes dessert.

1. List the jobs: for example research, check facts, write.
2. Give each agent a short instruction that says what it does and what it must not do.
3. Give each agent only the tools it needs. A researcher needs search; a writer does not.
4. Keep the number of agents small. Start with two or three.

Meera makes one worker per competitor, plus a writer. Workers get a search tool. The writer gets none.

You are done when: each agent’s job fits in one sentence.

Play 3: Pick a simple team shape

Goal: choose how the agents talk. Three shapes cover most beginner needs.

Start with manager and workers. It is the easiest to understand and the easiest to trace when something goes wrong.

You are done when: you can draw your team as boxes and arrows on paper.

Play 4: Write clear hand-off notes

Goal: stop agents from repeating or missing work. Anthropic reported that vague instructions to workers caused duplicated work, so each worker needs a goal, an output format, tool advice and clear limits.

1. Tell the worker its goal in one line.
2. Say exactly what to return, such as three bullet points and a source link.
3. Say what is out of bounds, such as “do not research other companies”.
4. Ask for a short answer, not a long essay, so the lead can read it fast.

Here is a small illustration of a hand-off note. The exact code depends on your tool; the idea stays the same.

task = {
  "goal": "Find current plans and prices for Company A",
  "return": "3 bullets + one source link each",
  "tools": ["web_search"],
  "do_not": "Do not research any other company",
  "max_steps": 8
}

In plain steps: it names the goal, the answer shape, the allowed tool, the boundary and a step limit.

You are done when: two workers given different companies never return the same facts.

Play 5: Set limits and watch the cost

Goal: keep the bill and the risk small. More agents means more tokens and more places to fail.

1. Set a maximum number of workers (for example, three).
2. Set a maximum number of steps per worker.
3. Give workers read-only tools unless they truly must change something. See our guardrails playbook.
4. Log every hand-off with a request ID so you can follow the trail. Our observability playbook shows how.
5. If a run may take long, save progress after each worker finishes, as in the long-running agents playbook.

Meera caps her team at three workers and eight steps each. She also prints the token count after every run (illustrative: one run used about four times the tokens of her single agent).

You are done when: a runaway run stops by itself.

Play 6: Test the team like a new hire

Goal: prove the team is better than one agent, not just fancier.

1. Make a small list of about 20 real examples of the task. Anthropic suggests starting with around 20 rather than waiting for a huge test set.
2. Run the single agent and the team on the same list.
3. Compare quality, time and cost side by side. Our evals playbook shows simple ways to score answers.
4. Keep the team only if it wins on what you care about.

Meera’s team finishes in about twelve minutes instead of forty (illustrative) and misses fewer prices. She keeps it.

You are done when: you have a side-by-side table that shows the team is worth the cost.

Do you need a team? A quick guide

Mistakes to avoid

• Too many agents. Start with two or three. Add more only when a test shows a gain.
• Vague hand-offs. Workers repeat each other. Fix it with the Play 4 note.
• No limits. A lead may start too many workers. Cap workers and steps.
• Shared context everywhere. If every agent needs to know everything, a team adds confusion. Use one agent.
• Never comparing with one agent. Always run the test in Play 6.

How to know it is working

• The team finishes faster than the single agent on the same task.
• Workers return different facts, not copies of each other.
• Cost per run stays inside the budget you set.
• When something breaks, your logs show which agent caused it.

Your checklist

• ☐ I wrote the task in one sentence.
• ☐ I know why one agent is not enough.
• ☐ Each agent has one job and only the tools it needs.
• ☐ I drew the team shape.
• ☐ Hand-off notes have goal, format, tools and limits.
• ☐ Worker and step limits are set.
• ☐ I compared the team with one agent on about 20 examples.

Frequently asked questions

Are multi-agent systems better than a single agent?

Not always. They help when work splits into independent pieces, like research. They cost more tokens and are harder to debug, so test first.

How many agents should I start with?

Two or three. One lead and one or two workers is enough to learn the pattern. Add more only when a test shows a clear gain.

Do agents need to use the same AI model?

No. Anthropic’s research system used a stronger model for the lead and a smaller one for workers. You can mix models to balance quality and cost.

What is the biggest risk?

Errors that pile up. One bad answer from a worker can mislead the lead. Checks, limits, logs and saved progress reduce this.

Closing: Meera’s Monday report

Meera now starts her Monday with a finished report. Three workers research three companies, a writer joins the results, and a short check catches wrong prices. She did not build it in one go. She started with one agent, found the real problem, added a small team and tested it.

Your next step: pick one task, run it with one agent, and write down where it struggles. Next in this learning path we look at how to keep AI cost and speed under control (model pricing, caching and batching), so follow the AI Playbooks category.

Sources: Anthropic, How we built our multi-agent research system; DEV Community beginner guide to multi-agent systems; upGrad multi-agent systems guide.