Google DeepMind has introduced computer use in Gemini 3.5 Flash, a new capability that lets the model perceive a screen and act on it — moving the cursor, clicking, typing, and navigating software the way a human would. This marks a shift from models that only generate text or images to ones that can operate tools directly.
The announcement frames this as a building block for AI agents that can handle everyday digital tasks, such as filling out forms, browsing applications, or managing files. By giving the model direct control over a computer interface, DeepMind aims to make automation more flexible than scripted workflows, which typically require bespoke integration for each program.
Because this is the only source reviewed, there is no independent confirmation of the details or claims. The blog post itself suggests that such capability comes with safety considerations, though specific safeguards are not described here. As with any early-stage agentic model, real-world reliability and oversight remain open questions.