How to give an AI agent its own computer
An agent that can only produce text is limited to advice. An agent with a computer can do the thing. Here is what that actually requires, and what changes once it has one.
Most AI tools answer in a chat box. You describe a job, they describe how they would do it, and you go and do it. That gap — between advice and work — is what an agent with its own computer closes. It opens the browser itself, fills the sheet itself, sends the mail itself.
"Giving an agent a computer" sounds like one thing but is really four, and products differ in which of them they actually give you.
The four things a computer has to be
1. Real, not simulated
A genuine operating system with a window manager, a real browser binary and a file system — not a text-only tool sandbox and not a headless scraper. This matters because most of the web assumes a real browser. Sites that block automation, portals that need a logged-in session, apps with drag-and-drop: these work when the agent is using the same software a person would, and break when it is not.
2. Persistent
This is the one most products skip, and the one that changes the most. If the machine resets between tasks, your agent wakes up with amnesia every session: logged out of everything, no files, nothing installed. Every task starts by re-establishing context that should already exist.
On a persistent machine the agent signs in to a supplier portal on Monday and is still signed in on Thursday. The spreadsheet it built last week is still there to add a column to. That compounds — the workspace gets more useful the longer it exists.
3. Watchable
If you cannot see what an agent did, you cannot trust it with anything that matters, and you will not delegate anything that matters. A summary at the end is not the same as watching. Summaries are written by the same system that did the work, so they inherit its mistakes: an agent that misread a page will confidently summarise the thing it misread.
Watching a live screen is a different kind of evidence. You see the page it actually opened and the number it actually typed.
4. Interruptible
Some steps a machine should not do alone — a login only you can complete, a CAPTCHA, a payment, a judgement call about tone. The useful design is not an agent that never hits those. It is one you can take over from mid-task: put your own hands on the keyboard, clear the obstacle, and hand it back without losing the work so far.
How the common approaches compare
| Approach | Real OS | Persists | You can watch | You can take over |
|---|---|---|---|---|
| Chat assistant with browsing | No | No | Partly — a list of steps | No |
| Hosted "agent mode" sandbox | Sometimes | Usually not | Sometimes | Rarely |
| Browser-automation scripts you write | Browser only | If you build it | If you build it | No |
| A VM you set up yourself | Yes | Yes | If you wire up a stream | Yes |
| MyClawn | Yes | Yes | Yes, live | Yes, mid-task |
The self-managed VM row is worth taking seriously. If you enjoy running infrastructure, you can build all of this: a cloud VM, a desktop environment, a remote-desktop gateway, an agent runtime, and the plumbing between them. It is a real option and it is cheaper in cash. It costs a weekend to build and ongoing attention to keep working, which is the trade.
What it costs
A dedicated cloud machine has an actual hardware cost, so anything offering one has to charge for it. With MyClawn the browser-based clone is free — it chats, remembers and learns how you work, it just has no machine of its own. A dedicated computer starts at $19/month, and every paid plan includes AI usage credits. The full tier table is on the homepage.
Worth knowing before you pay: spending is fail-closed — a request is refused before it can burn credits you do not have, rather than running up a bill. Usage is metered against rolling 5-hour, weekly and monthly windows, the same shape as a coding-assistant plan.
What it is genuinely good at
- Research that means opening pages. Twenty supplier sites, prices and lead times pulled into one sheet.
- Work behind a login. Dashboards and back-offices with no API, where a persistent session is the whole point.
- Outreach that ends in a sent message rather than a draft you still have to send.
- Long jobs. The machine is in the cloud, so it keeps going after you close your laptop.
Where it is the wrong tool
Being straight about this saves everyone time:
- One-off questions. If you want an answer, not a task performed, an ordinary chat assistant is faster and free.
- Work inside your own codebase. A coding agent that runs against your local repository fits that better than a separate machine that cannot see your files.
- High-volume structured extraction. If a real API exists for the data you want, use the API. A browser is the fallback for when there is no API, not a replacement for one.
- Anything requiring guaranteed unattended correctness. The agent can be wrong. The design assumes you are watching, which is why the screen is watchable and the takeover exists.
Related
- An AI agent that uses a browser — how the browsing actually works
- An AI agent that sends email — what sending as itself means
- Alternatives to ChatGPT's agent mode — how to pick
- Docs: what a cloud clone is