Blog ·
Most people start with one agent and give it more and more to do: the inbox, then the calendar, then a client's store, then the codebase. A month later it has access to everything, its instructions are three pages long, and when it does something odd, nobody can tell why.
The fix is the same one that works for people: give each agent one job and only the tools for that job.
Why several narrow agents beat one big one
- Smaller blast radius. If your research agent is tricked by a web page, it can't reach your inbox, because it was never signed in to it.
- Better results. Short, focused instructions and a small set of tools make agents more reliable than a long list of everything they might do.
- No login fights. Two agents sharing one browser profile sign each other out, overwrite each other's tabs and confuse each other's sessions.
- Easier debugging. When something goes wrong, you know which agent did it and you only have one job's logs to read.
- Clear costs. With a separate API key per agent, your provider's dashboard shows which job is expensive.
Three ways to keep agents apart
| Approach | Isolation | Effort | Good for |
|---|---|---|---|
| Separate user accounts on one machine | Weak: shared network, shared system | Low | Trying a second agent |
| Containers or virtual machines | Good, but usually no desktop | Medium to high | Terminal-only agents |
| A separate computer per agent | Strong: own files, browser and network | Low on a cloud computer | Agents that use apps and accounts |
Some agents also support several agents internally. OpenClaw, for example, can route different chats to separate agents with their own workspaces. That keeps their memory apart, but they still share one machine, one network and one set of installed skills. For agents signed in to different people's accounts, separate computers are the safer line.
A starter fleet
Here's a setup that works for a lot of people. Start with one or two and add the rest as you need them.
| Agent | Job | Signed in to | Talks to you on |
|---|---|---|---|
| inbox-agent | Sorts email, drafts replies for approval | A dedicated mailbox or delegated access | Telegram |
| research-agent | Compares suppliers, tracks competitors, writes briefs | Nothing private | Telegram |
| code-agent | Fixes tests and opens pull requests | One repository, via a scoped token | Discord |
| client-acme | Updates one client's store and reports back | That client's accounts only | The client's Discord channel |
OpenClaw is a natural fit for the inbox, Hermes Agent for research that repeats, and Claude Code, Codex or OpenCode for code. Our comparison of OpenClaw vs Hermes Agent and our guide to running coding agents in the cloud go deeper on each.
Five rules for a fleet of agents
- One job per agent. If you can't describe its job in one sentence, split it.
- Its own accounts. Each agent signs in only to what its job needs. Client agents never share logins with each other. See AI agents for small business for jobs worth giving them.
- Its own chat. A separate Telegram or Discord bot per agent, so you always know who's talking and can mute one without muting all.
- Its own API key and budget. Separate keys make spend visible per job and let you cut off one agent without stopping the rest.
- A clear name. Name the agent and its computer after the job, like
inbox-agentorclient-acme, not after the software.
When agents need to share work
Sometimes the research agent's findings should reach the agent writing the report. Keep these handoffs explicit and boring: a shared document, a repository both can read, or a file you pass along. Avoid letting one agent send instructions directly to another without review. If the first agent was tricked by something it read, that's how the trick spreads.
Running a fleet on Cubicles
Cubicles is built for one computer per job. Each computer is its own Ubuntu desktop with its own browser, files and logins, in a data centre rather than on your network.
- Add one in a few minutes. Choose New agent to set up OpenClaw, Hermes, Claude Code, Codex or OpenCode with its own model key and chat app, and it gets a computer of its own.
- See them all in one place. The Agents page lists every agent and whether it's ready, setting up or needs attention.
- Watch any of them. Open a computer from any browser to see its screen live.
- Share with a client. Send a view-only link to one computer's screen, so a client can watch their agent without seeing the others.
- Sleep the idle ones. Put a computer to sleep when its job is quiet. Its files stay for when you wake it.
Your first computer is free. Each additional one is its own flat monthly plan from $24, with no cap on how many you run. An account can have up to 20 agents.
Keeping an eye on a fleet
With more than a couple of agents, you'll want a routine. Have each agent send a short daily summary to its chat. Check your provider's usage dashboard weekly, per key. And if you live in an AI coding tool, connect it to your Cubicles account over MCP so it can list your agents, check their status, take a screenshot of a computer or set an agent up again. See giving your AI a cloud computer with MCP.
The bottom line
More agents isn't more chaos if each has one job, its own computer, its own accounts and its own budget. Start with the job that costs you the most time each week, give it an agent, and add the next one when the first is boring and reliable.
Frequently asked questions
Can I run multiple OpenClaw agents?
Yes. OpenClaw can run several agents behind one gateway with separate workspaces. For agents signed in to different accounts or clients, run each on its own computer so they don't share a network, files or installed skills.
Should multiple AI agents share one computer?
Only if they do related work with the same accounts. Agents that use different accounts, or work for different clients, are safer on separate computers.
How many agents can I run on Cubicles?
There's no cap on computers: each one is its own monthly plan, and your first is free. An account can have up to 20 agents.
Do several agents cost more in API usage than one?
Not necessarily. Usage depends on the work done, not the number of agents. Focused agents often use fewer tokens because their instructions and context are shorter.
How do I keep track of what each agent costs?
Give each agent its own API key with your model provider. The provider's usage dashboard then shows spend per key, and you can set a limit on each.