One Cursor per AI, and They Work in Parallel on Your Computer

Vercel Labs has added a project called agent-desktop. Its official one-line positioning is direct: agent-browser, for your whole computer. In other words, the agent capabilities previously provided for the browser have been pushed to the entire computer.

It lists four capabilities: provide Computer Use to any coding agent; support macOS, Windows, and Linux; run in the background while the user keeps working; and let multiple agents use the computer in parallel, each with its own cursor.

In the demo video, Notepad is open on the left with a Tokyo packing list, while Reminders and Calculator are on the right. It looks like very ordinary desktop operation.

Computer Use's value is not the packing list in the demo. Lincoln Powers made it clear in a reply: computer use is for all the background apps that never offered an API—that is, the software most people spend their days typing into to earn money. That judgment is closer to real demand than "AI clicks the mouse for you."

Project built by @mdlahfir, who just joined Vercel Labs. GitHub: https://github.com/vercel-labs/agent-desktop . Chris Tate's reply was direct: star, fork, tell your agent. Someone asked whether it can interact with agent-browser, and he answered Yes.

Multiple agents each getting a cursor sounds like multiplayer collaboration on the desktop. One commenter joked that agents can finally stop fighting over the mouse like siblings on a road trip. Another reply was more savage: agents on one desktop, each with their own cursor, and in the end the Mac will look like a group project Google Doc ten minutes before the deadline.

Jokes aside, the real question is boundaries. Jatin Garg asked: when one agent copies a password, what stops another agent from reading the clipboard? No answer appeared in the public replies. If this problem isn't solved, password managers, financial backends, and customer support systems are unlikely to be handed over to it with confidence.

There is also ownership. A Japanese commenter noted that in full-PC operation, more critical than environment differences is fixing "which tab is operated by whom"; background execution also needs to connect ownership and result confirmation before people dare to delegate. This observation is closer to real-world deployment than simply comparing platform support.

Canvas apps are another gap. Daniel Lenton asked what a snapshot looks like for an app like Figma that draws everything on a canvas. There was no answer in the replies. When Computer Use relies on windows and accessibility information, pure-canvas interfaces are usually harder to read.

There are already similar products on the market. Cua Driver claims to let agents like Claude Code and Codex use apps on a computer while the user keeps working, each with its own cursor, supporting macOS, Linux, and Windows, reading the system accessibility tree, and open source under MIT. Someone in the replies directly asked how agent-desktop compares with cua-driver; Chris did not elaborate in public replies.

So what is most certain about agent-desktop right now is its demo and positioning: cross-platform, background, multiple agents in parallel, independent cursors, and compatibility with agent-browser. What is most uncertain is the permission model, clipboard isolation, canvas app recognition, and who is in charge when multiple agents conflict.

The Tokyo packing list in the demo feels easy. A real workday is not this clean. To hand over the entire computer to an agent, the cursor is only the surface; scheduling and auditing are the lessons that still need to be filled in.