Today on @monad_dev's devrel livestream I showed what happens when an AI agent gets its own workspace, instead of stealing your screen.
Same agent. Same task. Watch the cursor.
The second time, I kept working while it worked. 👇
I'll post the build here as it goes. What works, what breaks, and the parts that never make it into a paper.
If you work on desktop agents: where does your execution layer break? That's the question I actually want answered right now.
Computer-use agents fail for a boring reason.
Not reasoning. Not vision.
They drive every Windows app through the same input primitive, and Windows does not agree that all windows are the same kind of window.
The paper is the research half. The other half is a plugin I'm building on top of it, running against real desktop apps.
Not public yet. I'm not going to pretend otherwise.
Spot on. "Don't Steal My Mouse" might just have to become the official tagline 🖱️🛑
The code and the research paper are dropping soon so you can finally get your desktop back. Stay tuned!
Today on @monad_dev's devrel livestream I showed what happens when an AI agent gets its own workspace, instead of stealing your screen.
Same agent. Same task. Watch the cursor.
The second time, I kept working while it worked. 👇
Exactly. The friction isn't just technical, it's cognitive. Once the agent shares the screen, it stops being a 'tool' and becomes a 'coworker' that you constantly have to watch. Invisible/parallel execution is the only way to solve that mental overhead. Building something similar for LaraCopilot?