Google unveils Project Mariner, an agent that operates a Chrome browser
The prototype scored 83.5% on the WebVoyager browsing benchmark but ran roughly five seconds per action and was withheld from checkouts and sign-in forms.
- Models & capabilities
- Notable
Google DeepMind showed Project Mariner, an experimental Chrome extension that let its Gemini 2.0 model take actions inside a web browser — clicking, scrolling, filling in forms and completing multi-step tasks such as shopping or booking travel from natural-language instructions. It launched the same day as the broader Gemini 2.0 announcement and Google’s Deep Research feature, forming part of a set of agent prototypes Google framed as its answer to browser- and computer-use agents shown earlier in 2024 by OpenAI and Anthropic.
The extension worked by taking screenshots of the active browser tab and sending them to Gemini in the cloud, which returned instructions for cursor movement and clicks; Google reported the agent achieved a state-of-the-art score of 83.5% on WebVoyager, a benchmark of real-world web browsing tasks. In practice the reported demonstrations were slow, with pauses of around five seconds between actions and the agent sometimes stopping to ask for clarification. Google restricted the agent to acting only within the currently active browser tab, a deliberate constraint intended to keep its actions visible to the user, and it was barred from completing purchases, entering payment or billing details, or accepting cookie prompts and terms of service — all treated as points requiring explicit human confirmation.
Access at launch was limited to a small group of trusted testers rather than the public. Google said it planned to broaden availability but gave no firm timeline. Project Mariner remained a standalone research prototype through 2025 before its capabilities were folded directly into Gemini rather than continuing as a separate product.