Watch an AI agent work — and take over when you want to

Most agents give you a summary of what they did. A watchable agent gives you the screen — while it is happening, with the keyboard one click away. Here is why that is the difference between a demo and something you can depend on.

There is a version of "AI agent" that is a black box: you describe a task, it disappears into a sandbox, and some time later a paragraph comes back telling you it is done. You did not see the page it opened, the field it filled in, the button it pressed, or the moment it misread something. You are asked to trust the paragraph.

That is acceptable when the cost of being wrong is a bad draft. It is not acceptable when the task touches a portal, a purchase, a message to a customer, or anything you would have to undo. For those, you want to watch. This page is about what that requires and what it changes. If you would rather see it than read about it, there is a real recorded run here — an agent sourcing suppliers on its own desktop — and you can take over from the last frame.

What a watchable agent needs

1. A screen that is actually the agent's

You cannot watch a sandbox that has no display, and you cannot usefully watch an agent that is driving your browser — you are already sitting there, and every click it makes is in your own accounts. A watchable agent needs a computer of its own: a real desktop with a real browser, separate from yours, that persists between tasks. Then there is something to look at, and looking at it costs you nothing.

2. A live stream, not a log

A list of tool calls is a translation. "Clicked element #submit" tells you less than seeing the form it was on. The honest artifact is the screen itself, streamed to your browser as it happens, so that what you see is what the site saw. On MyClawn the desktop is streamed live into the dashboard; the agent's current steps are shown beside it so you can follow its reasoning as well as its cursor.

3. A handover that is real

Watching is only half of it. The other half is being able to act: click into the stream and the keyboard and mouse are yours. Solve the CAPTCHA, type the password only you know, fix the form it got wrong, then hand it back and it carries on. Without the handover, watching is just anxiety with a picture. With it, the agent can stop at the parts it should never do alone, and the job still finishes.

4. A stop button that works

The human is priority one. If you can see it going wrong, you need to be able to end it — immediately, not after the task completes. An agent you can watch but not interrupt is a livestream of a mistake.

What watching changes

• You calibrate quickly. Watch an agent do a new kind of job three times and you know, concretely, what it gets right and where it hesitates. That is the actual basis for leaving it unattended later — not a benchmark, your own eyes on your own task.
• You catch the prompt-injection moment. The page that tells the agent to do something else looks, on screen, like exactly what it is. In a log it looks like a normal step.
• You can delegate more. Because you can step in for the password, the code and the "are you sure", the agent can take on work that a fully autonomous one has to refuse — anything behind a login, anything irreversible.
• "Done" becomes checkable. The machine keeps what it did — files, the sent folder, the browser history. You verify against the machine, not against the agent's summary of itself.

What watching does not solve

It does not make the agent smarter. A model that misreads a table still misreads it; you just see it happen. It does not remove the need for limits — the agent should still be set up to stop before paying or deleting, whether or not you are looking. And you will not watch everything: the point of the first few supervised runs is precisely so that you do not have to watch the hundredth. Treat the stream as the way you earn confidence, not a thing you must stare at.

Where this is the wrong tool

If the task is pure text — summarise, draft, translate — there is nothing to watch, and a chat box is the right shape. If you need a fully hands-off pipeline over public data at scale, a scripted scraper is cheaper than a desktop. A watchable agent earns its machine when the work is real, behind logins, occasionally needs you, and should keep going when you close your laptop.

See how the stream, the take-over and the interrupt are built in How it works, or watch a run and take over yourself.