Gemini Desktop Spotted Testing 'Computer Use' Feature for Direct Screen Control

Why Computer Use Is the New Battleground

AI agents that directly operate screens are territory already claimed first by Anthropic's Claude and OpenAI's ChatGPT. Instead of a person grabbing the mouse and clicking, the AI finds and clicks the spot itself, opening files and filling in content. The fact that Google is now belatedly testing this feature suggests it has closed some of the gap in the race to move Gemini from a chatbot that only answers questions into an agent that actually performs work on a user's behalf.

The Gemini desktop app has already introduced an interface called Gemini Spark through a feature that lets Gemini automatically plan travel itineraries. At the time, Gemini Spark pulled in Gmail reservation information and map data to build itineraries in Google Docs, and even filled out flight booking forms via Chrome on the user's behalf. The newly spotted Computer Use feature appears set to extend this Gemini Spark interface one step further.

How It's Expected to Work

According to TestingCatalog, Computer Use won't be a toggle switch in the navigation bar, but rather a "Task" option inside the prompt bar selector where image and video generation tools currently sit. Turning on this option would let users watch, through a mini-view of the active screen, exactly what actions Gemini is performing in real time. A practical control has also been prepared to prevent the computer from going to sleep during a task.

The most notable element is a safeguard called "Advanced Backup." Before Gemini touches a selected folder, this option automatically backs that folder up to Google Drive—giving ordinary users who don't rely on version control tools like Git a way to undo changes if something goes wrong.

A Double-Edged Sword

The very existence of this backup feature is itself a telling signal. Building in a way to revert changes beforehand amounts to Google acknowledging that Gemini could make unwanted file changes while carrying out a task. There are also practical concerns. If a selected folder runs into the hundreds of gigabytes, uploading it to Google Drive alone could require significant time and storage space, creating a bottleneck before the task even begins.

Item	Existing Claude/ChatGPT approach	Gemini test version
Activation location	Separate toggle/mode	'Task' option in prompt bar
Progress monitoring	Screenshots/logs	Mini-view of active screen
Undo safeguard	Relies on version control tools	Automatic Google Drive backup

Editor's Take

It's not hard to guess why Google held back on Computer Use and rolled it out last. Done well, this feature is convenient; done poorly, it leads directly to incidents that damage users' files. The fact that Google shipped it with a backup option front and center—after watching the trial and error Anthropic and OpenAI went through first, including incidents like clicking the wrong window or deleting unintended files—is itself a signal that Google is approaching this feature as a safety problem.

Having actually used agent features of this scale before, the conclusion is usually the same: you're impressed for the first few days, and then trust drops sharply the moment an unexpected file changes after you hand over an entire folder. The backup option is a minimal safety belt that softens the blow of that moment, not a complete solution. Teams looking to hand over entire project folders in the hundreds of gigabytes should, for now, limit Computer Use to low-risk tasks like document organization or repetitive form-filling. Folders containing important codebases or customer data should probably wait until finer-grained permission controls arrive after the official launch.

In the coming weeks, this feature will likely expand beyond the trusted-tester pool to a broader user base. But first impressions will hinge on how Google refines permission scope, recovery procedures, and handling of large folders.