How to stop an AI agent mid-task

Most agent tools give you one way to stop something: close the tab. That works, but it tells you nothing about whether the agent actually stopped, and it throws away everything it was in the middle of. A real interrupt is a different, smaller promise — and it's worth being precise about what it actually guarantees.

Closing a browser tab isn't stopping an agent — it's stopping watching one. If the work runs on your own machine, the process usually dies with the tab, which is fine until the one time it doesn't: a background job, a scheduled run, a task that kept a connection open somewhere you can't see. If the work runs on someone else's server, closing the tab does nothing at all — it keeps going, and you find out what it did later. Either way, "stop" needs to be a real command the agent receives and acts on, not a euphemism for looking away.

What a real interrupt has to do

1. Reach a machine that is actually still running

You can only interrupt a task that has a place to be interrupted. On MyClawn the clone works on its own persistent cloud desktop — see what giving an agent its own computer involves — so a stop request has somewhere concrete to land: a real session, on a real machine, doing one identifiable thing at a time.

2. Answer honestly, not optimistically

The dangerous version of a stop button is one that always says "stopped" whether or not it actually reached the machine. MyClawn's interrupt call to the clone's desktop either confirms the stop or reports plainly that it couldn't reach the machine — it does not report success it didn't verify. Telling you the agent stopped when it didn't is worse than not having a stop button at all, because you'd stop watching.

3. Hand you the screen, not just a cancelled request

Stopping is rarely the whole goal — usually you stop because the agent misread the task and you want to fix something yourself. The desktop it was working on is still there afterward: the browser tab it had open, the form half filled in, the file it was editing. You can watch what it did live, or reach in and take the controls the moment something looks off, correct the one step only you can correct, and hand the task back rather than starting over.

Why this is the whole point, not a safety feature bolted on

An agent that can't be interrupted mid-task is an agent you have to fully trust before you let it start — because once it's running, your only real option is to wait and see. That's a much higher bar than most tasks deserve, and it's why a lot of agent work still happens one small, watched step at a time instead of being handed off. The alternative isn't more trust — it's not needing as much of it, because the human stays able to stop or take over at any point, not just at the end. That's the design question behind the whole product, not one setting in it: the screen is always there, and nothing the agent does is ever off-limits to you.

What we don't claim

An interrupt stops the agent's next action — it does not undo the ones already taken. If it already sent the email or submitted the form, stopping it a second later doesn't unsend it. Watch anything irreversible happen — and afterwards, every mail it sent is in the sent-mail view — rather than treating "I can stop it" as a substitute for watching in the first place.

Related

• Watch an AI agent work — and take over when you want to
• Giving an AI agent its own computer
• An AI agent that works while you sleep