Google adds built-in computer use to Gemini 3.5 Flash

Google announced that “computer use” — the ability for a model to see an interface, reason about it and take actions as part of an agent workflow — is now integrated as a native tool in Gemini 3.5 Flash. The feature is aimed at developers and enterprises building agents that can operate across browser, mobile and desktop environments.

What happened

Google’s post, dated June 24 and updated June 25, says computer use is moving from a standalone Gemini 2.5 computer-use model into Gemini 3.5 Flash itself. According to the company, Gemini already supported function calling and tools such as Search and Maps grounding; the new step adds a built-in computer-use tool to the Flash model.

Google frames the capability around agents that can see, reason and act on interfaces. Developer documentation and a reference implementation show the feature is intended for workflows where an agent interacts with screens, applications or websites, not only text. Initial access is through the Gemini API and Gemini Enterprise Agent Platform.

Why it matters

The announcement matters because the next phase of AI agents is not just answering questions or generating instructions. To automate real work, systems need to operate existing tools: web forms, internal dashboards, office apps, testing systems, CRMs, tickets, reports or development environments. Integrating computer use into a general model such as Gemini 3.5 Flash narrows the gap between an agent prototype and a workflow that can touch everyday software.

For enterprises, the appeal is long and repetitive work: continuous software testing, administrative tasks across apps, reviewing information in legacy systems or coordinating steps that today require humans to move between screens. For developers, the signal is that Google wants Gemini to compete not only as a conversational model, but as a foundation for agents that execute actions.

What changes for users and technical teams

In practice, this does not mean every agent should get unrestricted computer access. What changes is the product layer: instead of stitching a model to external screen-control tooling by hand, teams can test an integrated Gemini 3.5 Flash capability with official APIs, documentation and examples.

That could accelerate automation pilots, especially for teams already using Gemini or Google Cloud. It also raises competitive pressure on other agent providers because the graphical interface — not only APIs — is still where many enterprise processes live.

Product and automation reading

From a product perspective, the important point is convergence: a fast model, built-in tools, grounding and the ability to act on interfaces. That combination brings agents closer to real processes, but it also requires careful design. Automating a screen is not the same as understanding the business process behind that screen; without permissions, observability, audit trails and limits, an agent can execute the right-looking step in the wrong place.

Google acknowledges part of that risk. The announcement mentions targeted adversarial training against prompt injection and two optional enterprise safeguards: requiring explicit user confirmation for sensitive or irreversible actions, and automatically stopping tasks if an indirect prompt injection is identified. The company also recommends sandboxing, human verification and strict access controls.

What remains unclear

Independent evidence is still needed on reliability in real enterprise environments, costs, usage limits, compatibility with complex applications and behavior around dynamic pages or sensitive data. What is confirmed is the availability and architectural change: computer use is now part of Gemini 3.5 Flash as an integrated tool, with official documentation for developers to start building.

Sources consulted

Google DeepMind / Google Blog official announcement: Read More AI for Developers computer-use documentation: Read More Gemini reference implementation: Read More by Nova Rivera — Product and automation perspective.