Google DeepMind has embedded computer use as a native, built-in tool in Gemini 3.5 Flash — marking a significant step forward for agentic AI. Previously offered only as a separate Gemini 2.5 computer use model, the capability is now integrated directly into the main Flash model, making it accessible to a much broader set of developers.
What Computer Use Enables
Gemini 3.5 Flash can already leverage function calling and built-in tools like Search and Maps grounding. With computer use baked in, developers can build custom agents that:
- See on-screen content across browser, mobile, and desktop environments
- Reason about multi-step tasks and user interfaces
- Take action autonomously within those environments
Target use cases include continuous software testing, accessibility auditing, and knowledge work automation across professional applications — tasks that require sustained, long-horizon reasoning.
Access and Availability
The feature is available now through two channels:
- Gemini API — for developers building custom agent workflows
- Gemini Enterprise Agent Platform — for enterprise-scale automation deployments
A live demo environment is also hosted by Browserbase for those who want to test capabilities before committing to an integration.
Safety Architecture
Operating agents in live environments introduces real risks, particularly around prompt injection — where malicious content in the environment attempts to hijack agent behavior. DeepMind's response is a layered, "defense-in-depth" approach:
- Targeted adversarial training to harden the model against injection attacks
- An optional safeguard requiring explicit user confirmation before sensitive or irreversible actions
- Automatic task termination if an indirect prompt injection is detected
DeepMind also recommends pairing these features with secure sandboxing, human-in-the-loop verification, and strict access controls. Full guidance is available in the team's best practices documentation.
Enterprise Signal
DeepMind reports that early customers are already extracting value from the capability, though specific case studies weren't detailed in the announcement. The integration of computer use into a broadly available, production-grade model — rather than a specialized research preview — signals that agentic automation is moving from experiment to infrastructure.
For startups building on top of this capability, a polished AI product website can be critical for communicating the value of agent-driven workflows to potential customers and partners.
