At Novastart (https://novastart.com/) we built an infinite canvas OS: every desktop app lives on one endless canvas, and workspaces restore exactly as you left them. Then we let AI agents work on that same canvas, each with its own mouse and keyboard.
Three things we learned
- One lead, many helpers. A lead agent drafts a plan with one task per helper and opens a window for each. Up to 10 agents per task stay manageable because each has a visible place on the canvas.
- Humans must be able to step in. Pause freezes everything. You can type a correction mid-run and the agents adjust without restarting, or just take a window back.
- Done means checked. Agents report success they didn't achieve. A separate check that looks at a fresh screenshot before marking a task complete was our single biggest reliability win.
The same canvas is multiplayer, so teammates share the workspace (own cursors, chat, each other's apps) instead of sharing a screen.
What patterns are you using to keep computer-use agents honest?
