How to give AI agents their own computer (cloud VMs and computer use)

By default, an AI agent that uses a computer is using yours. It grabs your mouse, it stops when you close the lid, and everything on your machine is in reach. If you want agents that work like employees, they need somewhere to work that is not your laptop.

What "their own computer" means

It means a separate machine, usually a virtual machine (VM) in the cloud, that belongs to the agent. The agent logs in, opens apps, runs code and keeps its files there. You can watch it, step in or leave it alone.

The Cua docs describe a sandbox as an isolated machine that can be a container or a VM, created and deleted on demand. With the optional cua-spacesd service running inside it, an agent can run shell commands, move files, take screenshots, and send input to the windows on that machine.

Other tools take the same idea inside a product. Devin says it spins up its own Windows, Mac and Linux virtual machines to build, test in a browser and record itself clicking through the result. Codex Cloud runs coding tasks on OpenAI-managed computers.

Why it beats using your own machine

• It keeps working when you stop. Your laptop can be closed.
• You can run many at once. Each agent gets its own machine, so one never fights another for the screen.
• Mistakes stay contained. A bad run happens on a disposable computer. Cua notes that deleting a sandbox discards its state, so save anything you need first.
• It fits real roles. A QA agent, an assistant, an ad worker or an engineer each gets a desk of their own.

How to set it up

Exact steps depend on your tools, so check each product's current docs. The shape is always the same:

1. Create a machine. A cloud VM, or a sandbox on a computer you own. Cua's CLI can create sandboxes locally or in your own AWS or Google Cloud account, and connect machines you already own.
2. Pick the operating system the work needs. If you are testing an iOS app, you need a Mac. Otherwise Linux is the cheapest to run.
3. Install a computer-use driver on it. Cua Driver lets an agent use apps in the background on macOS, Windows and Linux.
4. Connect your agent. Cua says it can add Spaces to agents such as Claude Code, Codex and Cursor.
5. Give it a job with a finish line. Not "do marketing" but "review these ads and list five changes".
6. Watch a few runs, then stop watching. Use the live view to catch mistakes early, then rely on reports.

How we do it at SupaKid Studios

Tomi gave his agents their own computers to scale his own company to what he describes as 50 employees. In his words, a Cua driver is a faster, better version of computer use, and it works with any AI agent. He sets up VMs, which are computers that live in the cloud, and has the agents install their own Cua drivers on them.

He says he can run 50 or 100 agents, each on its own computer. Some do different tasks. Some do QA. One is an assistant employee. For a solo founder, his example is a marketing agent: give it a cloud computer with computer use and it can work on ads, look at customer research and do so without your computer being open.

The agent can be hooked up to Claude Code, Codex, Hermes or whichever agent you use. And he can watch all of the machines natively from his Mac, scrolling between the VMs his agents have access to. He does not surrender his own computer to get the work done, and the agents persist on their machines, so they work like remote employees.

This is also how QA works in his factory: Devin tests on its own machine. See AI code review and testing agents, and for the larger setup, what an AI software factory is.

What goes wrong

• Giving it your logins. An agent with a computer can reach whatever you sign in to. Create separate accounts with only the access the job needs.
• No finish line. An agent with a desktop and no task wanders.
• Forgetting the bill. Machines that run all day cost money. Stop the ones you are not using.
• No human check on outputs. Anything that goes out to customers or the public should be approved first.
• Lost state. If a sandbox is deleted, so is its work. Save results to a repo or shared storage.

Want a team of these without building it?

Setting up the machines, logins and guardrails is the dull part, and it is where most attempts stall. SupaKid Studios sets up and runs agents like this as part of a managed AI agent team, and you approve anything that goes out. If you want to see what that would look like for your work, book a 30-minute call.