Google's Project Mariner AI Redefines Web Browsing with Gemini 2.0

Google launched Project Mariner on December 11, a Chrome extension that lets an AI agent take over a browser tab and perform tasks autonomously. Built on the newly announced Gemini 2.0, Mariner can read web pages, fill out forms, click through multi-step workflows, and navigate sites—all while the user watches from the same browser window. Google is calling it an experimental release, available through a waitlist inside the company’s Trusted Tester program.

The launch puts Google directly into the emerging AI agent market, alongside Anthropic’s Computer Use feature for Claude, which lets the model control a desktop computer, and OpenAI’s reported Operator agent, which is expected to launch in early 2025. The common idea across all three is that large language models have matured past the point where the only useful output is text. They can now take action.

Mariner works by integrating with a Chrome tab through the existing extension API. The user types a command—”find me a flight from San Francisco to Denver under $300”—and the agent takes over the tab, searching Kayak or Google Flights, filling in dates and destinations, and presenting options. The user can take back control at any time by clicking or typing, and the agent pauses its automation when the tab is not visible. Google has also restricted Mariner from submitting payments or agreeing to terms of service without explicit human confirmation.

Gemini 2.0, the model powering Mariner, scored 83.5% on the WebVoyager benchmark, which measures an agent’s ability to complete real-world web tasks. The previous state of the art was 56%.

The WebVoyager score is eye-catching, but the benchmark measures performance on curated tasks in controlled environments. Real web browsing is messier. Pop-ups, cookie consent banners, CAPTCHAs, unexpected redirects, and sites that load content dynamically with JavaScript all present failure modes that do not show up on a clean test suite. Mariner’s actual reliability on the open web is unknown.

Google is also positioning Mariner as a research prototype rather than a consumer product, which gives the company room to iterate before the stakes are high. But the broader signal is clear: Google sees AI agents as the next frontier after search summaries. If AI Overviews are changing how people find information, Mariner is about what happens after they find it. The agent does not just answer your question about flights. It books the ticket.

The timing of the Gemini 2.0 and Mariner announcements, landing in the same week as Google’s GenCast weather model and its expanded AI Overviews, suggests a company trying to reclaim the narrative after a year in which OpenAI and Anthropic dominated the conversation around frontier AI. Whether Mariner works well enough to shift that narrative will depend on what the waitlisted testers find.