Google is advancing its Agent feature within Gemini, targeting users who need help with complex web tasks that go beyond simple chat. The Agent prototype is already referenced in new interface banners, describing it as an early-stage research tool capable of browsing, interacting with websites, using tools and services, and executing user instructions across sessions. Notably, it can maintain sessions, keeping users logged in across multiple actions, so it could be helpful for workflows involving multiple steps online, like research, booking, or data gathering.
BREAKING 🚨: Google is preparing Gemini Agent prototype to perform your tasks on the web!
— TestingCatalog News 🗞 (@testingcatalog) October 15, 2025
It is likely to be powered by the recently released Computer User model and will be able to operate a browser and perform deep research.
With great power comes responsibility 🤖 pic.twitter.com/LcXnphOVEL
From the banner details, Google positions Agent as powerful but still experimental, emphasizing user responsibility and caution around sensitive tasks or credentials. Users are warned not to share login or payment information directly in chat, and to supervise Agent closely, especially when handling data or automating actions. This messaging is in line with Google’s cautious approach to deploying agentic automation, particularly as it relates to privacy and security.
While Agent is not yet available to the public, its presence in labs and recent UI updates suggest a release could happen soon, likely to early testers and possibly rolling out to broader Gemini users later. The absence of a visible model selector in Agent mode hints that a dedicated backend, potentially Gemini 3 or a special browser-optimized model, will handle these tasks, further separating Agent from existing Gemini Pro or Flash experiences.
This move would reinforce Google’s broader product direction toward integrated, task-based AI that can automate, monitor, and adapt to real user workflows across web and app environments, echoing trends seen in OpenAI’s and Anthropic’s agentic features. The ability to automate complex tasks could become a key differentiator for business users and power users seeking efficiency gains in their daily routines, while also requiring clear guardrails and user education for safe deployment. The Agent feature, if successful, could push Gemini into new domains of productivity and automation not yet accessible to most consumers.
