SNACK three-line summary
- Google has officially added Computer Use to Gemini 3.5 Flash. By looking at screenshots and understanding browser, mobile, and desktop screens, the model can suggest UI actions such as mouse clicks and keyboard input.
- This change is less about improving simple chat answers and more about agents moving into the stage of actually handling apps and the web. Google also presented example tasks such as filling out forms, testing websites, and researching across multiple sites.
- It would be an exaggeration to read this as fully unattended automation. Google emphasized that user confirmation may be needed for risky or hard-to-reverse actions, and recommended using sandboxes, human review, and access controls together.
Snackgirls editor note
AIKO: “It feels like AI is moving beyond just writing text and into the stage of looking at a screen and suggesting the order for pressing buttons.”
Red: “But the important point here is not the fantasy that it magically does everything on its own. It is how far to automate, and where to insert human confirmation. In real work, that boundary matters more.”
What changed
Google announced on June 24 that it had added Computer Use to Gemini 3.5 Flash as a built-in tool. The feature works by letting the model view a screen through screenshots and suggest the next UI actions, such as mouse clicks and keyboard input.
According to the official documentation, the supported environments are browser, mobile, and desktop. Google explained that developers and businesses can use this to build agents for tasks such as automating form input, testing web applications, and carrying out research across multiple sites.
Why it matters
The key point of this announcement is that AI is expanding beyond the stage of producing answers or code and moving toward an operational tool that can read real screens and suggest sequences of action. Put simply, it is becoming less like an AI assistant that only writes text well and more like a work helper holding the remote control.
Google said Computer Use is especially useful for long-step automation and enterprise work. The official documentation also says it can handle browser, mobile, and desktop from one model axis, and attach intent to each action to explain why it is trying to perform that action.
What general readers may misunderstand
However, it would be an exaggeration to see this feature and assume that the Gemini app can freely control every website and program starting today. The official documentation says developers need to implement the client-side execution environment separately, and the current focus is on building agents for developers and enterprises.
In other words, this announcement is less a consumer shortcut-button news item and more a signal that agent development tools have expanded by one step. For general readers, it becomes a new reference point for watching how far service automation expands into on-screen workflows.
What to watch carefully now
Google also emphasized safety measures. The official announcement includes user confirmation for sensitive or hard-to-reverse actions, stopping work when indirect prompt injection is detected, and documentation recommendations for sandboxes, human-in-the-loop review, and strict access controls.
So from a practical work perspective, designing where the system should stop and who should confirm it is just as important as asking “what can be automated?” This announcement clearly shows that expanded agent capabilities need to come with safety operation rules at the same time.
Sources and checked date · Published 2026-06-24 / checked 2026-06-25T01:36:46+00:00
Sources
Leave a comment