For years, AI coding assistants have been the world's most talented backseat drivers: brilliant at suggesting where to turn, totally unable to touch the wheel. Cursor just handed them the keys. Its agents can now click and type their way around your desktop, which is either the future of QA or the opening scene of a very polite techno-thriller.
From the Browser to Your Whole Desktop
On October 23, Cursor announced Computer Use, a feature that lets its coding agents "click and type across any app on your computer, not just the browser." The headline example is Apple's iOS Simulator: an agent can build a feature, then actually open the simulator and test what it just built. You can ask for it by @-mentioning Computer Use in a prompt, or let the agent switch it on by itself when a task needs it.
This builds on earlier work. Cursor's cloud agents already ran in isolated virtual machines with their own desktops. What's new is that the capability now reaches the developer's own machine. On macOS it runs through a small helper app the CLI manages, and self-hosted Macs need explicit operating-system permissions. Cursor announced no accompanying pricing change.
Trust, but Keep It on Manual Mode
The appeal is obvious. A solo developer gets a tireless tester who doesn't complain about clicking through the same signup flow forty times. Closing the loop between "write the code" and "see if the code works" is exactly where agents have been weakest, and giving them eyes and hands on real interfaces is a meaningful step.
The catch is just as obvious. An agent that can drive any app can also drive the wrong app, with your logged-in sessions, your files, and your credentials sitting right there. Early reactions reflect that: some developers love the QA potential, while others plan to keep it strictly on manual for now. Our take is that permission scoping matters more than ever. Grant access only to the helper, run it somewhere disposable when you can, and treat "the agent decided on its own" as something you audit, not something you assume is fine.
The backseat driver now has the wheel, so maybe don't leave your banking app open in the next window.
If your team is weighing agentic AI tools against real-world risk, talk to us about rolling out AI with sensible guardrails so the robots test your app without wandering into your admin panel.
Source: RuntimeWire