One agent or several? How to split work between AI agents

One AI agent doing every job gets worse at all of them, because every job's rules sit in front of it every time. Split off one job at a time. Give each agent a short card it reads first: what it owns, what it never touches and whose that is, who it hands work to, and what counts as done. The quickest way to get a team like this is HostNest, set up for you in about 10 minutes; or use the steps below with agents of your own.

Setting up your very first agent? Five lines are enough to start: give an agent one job in five lines of CLAUDE.md. This page is for the second agent, and every one after it.

How do I get a team of agents without setting each one up?

The quickest way is HostNest, a team of Claude agents. You sign up yourself, and the rest comes ready: your own server in the EU (a computer that's always on), the agent, and its wiring. It takes about 10 minutes, most of it the server being built:

1. Account: your email (we send you a 6-digit code) or Google.
2. Free trial: 7 days free.
3. Telegram: a few taps create your bot. The bot is yours.
4. Claude: while your server is built, you read four safety rules. Then you sign in to your paid Claude plan (Pro, Max, Ultra or Team) and paste back one code.
5. Live: open your first agent in Telegram.

Then you add teammates from the dashboard, each with its own Telegram chat and its own job. Send each one its card (below) in its own chat, and ask it to keep it in its notes. Teammates can hand each other work through a shared mailbox, which stays off until you say yes. We recommend up to 4 teammates alongside your main agent.

After that, the upkeep is handled for you. Updates install themselves overnight, a failed one rolls back on its own, and the server is backed up daily. Get your team on HostNest

One HostNest user added a second agent as an electronics tutor. Troubleshooting a breadboard with him, it found the 12 V at the input that had killed a board.

Do I need to code? No. You talk to your agents in Telegram. How long does it take? About 10 minutes, most of it the server being built. Will it do things without asking? It checks with you before anything goes out or costs money, unless you tell it it doesn't need to. Is my stuff safe? It runs on your own private server in the EU, with your own Claude login, and it's backed up daily.

Should one AI agent do every job?

No, not once you have more than one kind of work. An agent works with whatever is in front of it. A rules file written for every job puts the wrong rules in front of it most of the time. Chroma's 2025 study of 18 models found that "model performance consistently degrades with increasing input length", even on simple tasks. Anthropic's best practices for Claude Code say the same about a rules file that tries to cover too much. When an agent keeps ignoring a line, one reason they give is that "the file has grown long enough that the line gets lost."

When should you split off a second agent?

When one job keeps getting rules meant for another, or you keep correcting the same job. Split off the job you repeat most, and only once you can write its card. More agents cost more, so add them one at a time.

What goes on each agent's card?

Five lines, in this order:

1. Its job, in one line.
2. What it owns: the files, outputs and decisions that are its own.
3. What it never touches, and whose that is.
4. Who it hands work to, and how.
5. What counts as done: the file, link or number that proves it.

What does the card look like?

Copy this, and fill in each line for your agent:

# Your job
<one line: what you do>

# You own
- <the files, outputs and decisions that are yours>

# You never touch
- <what's not yours, and whose it is>

# You hand work to
- <who gets what from you, and how you send it>

# Done means
- Never report done on its own. Report it with the file path, link or number that proves it.

Put it at the top of the agent's own rules file, the one it reads at the start of every session. In Claude Code that's the CLAUDE.md in its folder.

How do two agents stop doing the same work?

With lines 3 and 4. Each card names what the other agent owns, so one agent's "never touches" is the other's "owns":

# The writer
You own: the story.
You never touch: how it looks and sounds. That's the editor's.

# The editor
You own: how it looks and sounds.
You never touch: the story. That's the writer's.

A builder ran seven coding agents for 78 hours. They wrote that the agents "collided on shared files". They also wrote that the setup "works surprisingly well when responsibilities, communication, review gates, and decision history are clearly defined." Anthropic's research agents each get "an objective, an output format, guidance on the tools and sources to use, and clear task boundaries."

How do agents hand work to each other?

Through something both can read: a shared folder, a task list, or a message from one to the other. Say which on line 4, so the work doesn't wait for you to carry it across.