Google DeepMind has added computer-use capability directly into Gemini 3.5 Flash, giving developers a built-in way to create agents that can interact with software across browser, desktop and mobile environments.
The update moves the feature from a standalone model into the company’s main Flash offering. Google said the change is designed to improve performance for agentic tasks, especially workflows that stretch across multiple steps and require an AI system to observe a screen, reason about what it sees and take action.
According to Google, Gemini already performs well at function calling and at using integrated tools such as Search and Maps grounding. The addition of computer use expands that toolkit, allowing developers to build custom agents that can operate across platforms rather than within a single app or website.
Google said the feature is aimed at longer-running automation work, including continuous software testing and knowledge work in business settings. The company described Gemini 3.5 Flash as its strongest option so far for computer-use tasks.
The feature is now available through the Gemini API and through Google’s Gemini Enterprise Agent Platform. Google also pointed developers to a reference implementation and documentation for those who want to begin testing the system.
To help demonstrate the capability, Google highlighted example uses such as analyzing the Gemini app and producing a categorized list of features, as well as reviewing documentation for accessibility issues. Those examples suggest the company is positioning computer use not only as a general automation tool but also as a support for product analysis and internal operations.
Google acknowledged that agents operating in live environments face prompt injection risks, where malicious instructions can be hidden inside the content they process. To address that, the company said it used targeted adversarial training for the Gemini 3.5 Flash computer-use feature.
It is also offering two optional enterprise safeguards. One requires explicit user confirmation before sensitive or irreversible actions can go forward. The other can stop a task automatically if the system detects an indirect prompt injection.
Google said developers should pair these protections with secure sandboxing, human review, and strict access controls. The company framed the approach as defense in depth, meaning multiple layers of safeguards rather than reliance on a single control.
Google said some customers are already using computer use in production-like workflows and seeing value from it, though it did not provide detailed case studies in the announcement. The company included quotes from partners including Browserbase, Browser Use, and UiPath, signaling interest from firms focused on browser automation and enterprise workflow software.
Developers can try the capability in a demo environment hosted by Browserbase and then move to the reference implementation and platform documentation for building their own agents.
The release reflects Google’s broader push to make Gemini a more capable agent platform, not just a general-purpose model. By folding computer use into Gemini 3.5 Flash itself, Google is making the feature easier to adopt for teams that want AI systems to interact with real software interfaces as part of larger automated workflows.