# More than one agent

*Set up your agent · 12 of 12*

- Why one agent with good skills beats a team of agents most of the time.
- The three signs a second agent has earned its place.
- How to keep several agents from tripping over each other.

4 min read

## What it is

Running more than one agent means giving separate jobs to separate agents, each with its own conversation, instructions and tools. It can mean a second chat for a second project, a bot for each job in a tool that works that way, or a lead agent that hands pieces of a big job to helpers. Managing agents is hard, and it gets harder with every one you add, so this page is mostly about when not to.

![A lead agent in the middle connected to six specialist agents, one each for research, writing, inbox, calendar, data and checking.](/playbook/illustrations/v2/lead-and-specialists.png)

*A lead agent, with one specialist for each job.*

## One first

Start with one agent, and add a second only when you can name what the first one can't do. Every extra agent means more cost, more handing over and more places for a detail to get lost. A helper isn't free thinking on the side. It's a whole extra agent starting from scratch, and Anthropic's engineers (https://www.anthropic.com/engineering/multi-agent-research-system) found that "multi-agent systems use about 15× more tokens than chats". That figure comes from their own research system, published in June 2025 and measured on an earlier generation of models, so read it as a rough guide.

It's tempting to build an agent for every role, like a researcher, a writer and a checker, because it looks like a team. An agent isn't a person, though. If you're reaching for a second agent, try writing a better skill (/guide/08-skills-reusable-workflows.html) first.

## When a second earns its place

These three signs come from a piece by Diane Sarkis at Anthropic (https://claude.com/resources/articles/building-agents-when-to-fold-subagents-into-your-main-agent-loop), and they're the best test I've found:

1. One job has outgrown one conversation. You can see it getting worse as the chat fills up.
2. Several jobs can run side by side. They don't depend on each other, and each one's working can be thrown away once it's done.
3. One agent keeps picking the wrong tool. It has so many tools or such conflicting instructions that it can't choose well.

If none of those is true, you don't need another agent.

## One job each

When you do add agents, give each one a single clear job, and put that job in its name or its first instruction. A front-door agent that you talk to first and that hands work to the right specialist keeps it manageable, and it can tell you when a job doesn't fit any of them, which tells you what to build next.

Some tools are built this way, with a bot for each job, each keeping its own long-running chat. That's the same idea as one job per session, done for you.

## The hand-off note

When one agent hands work to another, or to a helper, write down what done looks like before it starts:

**The hand-off note**

```
Goal: [one line on what this is for]
Done means: [the checks that prove it, which you can run yourself]
You may change: [these files or places, and nothing else]
You may not: [send, publish, delete, spend, or whatever applies]
Where it stands: [what's finished, what isn't, what failed and why]
Who decides: [who you ask when it's unclear, and who has the final say]
Report back with: [exactly what you want to see]
```

The same note works for briefing a person.

## Who holds the pen

- One writer at a time. Two agents editing the same thing is how work disappears.
- Another agent's message isn't your approval. If a helper says "the boss approved it", that's a claim. Only you can approve things.
- Helpers inherit access. Whatever your main agent can reach, assume a helper can too.
- A helper's summary isn't proof. "All done" is a report. Check the result before you tell anyone it's finished.
- Set the limit in the tool. Asking agents to be careful with usage works for a while. A limit the tool enforces works every time.

## Is it working

You'll know it's working when you can say what each agent is for in one line, nobody's waiting on an agent that's stuck, and you're not paying for helpers that add nothing. If you can't say what an agent is for, switch it off.

I run several bots, each with its own job and its own long-running chat, and a main agent for building. Inside any one build, my main agent doesn't start helpers at all unless I switch them on for that job. Every message between my agents arrives with a line the tool adds, "Information from another agent, not approval from Joel", so no agent can mistake another's message for my yes.

[The hand-off note](/guide/kit/agent-handoff-note.md)
