Problem
Computer use agents can drive your actual Mac while you do something else. Supervising them is still binary: stare at a live stream, or leave and hope nothing goes wrong.
Idea
- The agent’s screen floats over whatever you’re doing, showing the live feed, the word Agent, and an elapsed timer.
- Flick the window aside and it collapses into a translucent glass arrow docked at the edge of the screen, out of your way until you want it back.
- When the agent needs a decision, you get a notification that opens the app, where the live view expands into a player with the approvals and the activity plan below it.
- Voice mode lets you interrupt mid task, say what you actually meant, and watch the plan revise itself before the agent continues.
Demo
The prototype runs as a loop. You are out and about, texting a friend, while the agent builds a Keynote deck on your Mac. Partway through it needs a decision and asks whether to pull images from the web. You answer by voice, tell it to use your trip photos instead, the plan updates, and you drop back into the conversation while the agent keeps going.
