Project Mariner

A prototype browser agent

Project Mariner is a Google research prototype: an agent that does its work inside a web browser. Google announced it on 11 December 2024 alongside Gemini 2.0, and the prototype is built on that model.^[1] Its framing places it in the "agentic" line of work — systems that do not only answer questions but take actions on a user's behalf.

How it works

Mariner runs as an experimental Chrome extension. Rather than calling a site's internal interface or relying on page code, it completes tasks by understanding what is on the screen — the same rendered page a person would look at — and then acting on it.^[1] That approach is what makes it general: a browser agent that reads the screen instead of depending on a fixed set of site integrations can, in principle, be pointed at sites it was never specifically built for.

Google reported 83.5% on the WebVoyager benchmark, which tests real-world web tasks, describing the result as state of the art and noting that the system worked as a single-agent setup.^[1] The single-agent detail matters: some competing approaches split a job among several cooperating models, so one agent reaching that score is a simpler design.

Guardrails

An agent that clicks and types in a live browser can do real damage if it goes wrong, so the prototype was deliberately limited. It could only type, scroll or click within the active tab, and before sensitive actions — a purchase, for example — it asked the user for final confirmation.^[1] The design assumption is that the human stays in the loop at the moments that matter, while routine navigation and form filling are delegated.

From prototype to developer API

At Google I/O in May 2025, Google said Project Mariner's computer-use capabilities were coming to the Gemini API and Vertex AI, naming partners including Automation Anywhere, UiPath and Browserbase.^[3] The move turned an internal research demonstration into something developers could build on, and it brought in vendors from process automation and browser infrastructure whose customers already automate web workflows.

A concrete artifact followed on 7 October 2025, when Google released the gemini-2.5-computer-use-preview-10-2025 model in the Gemini API; that preview model was retired on 28 July 2026.^[2] Preview releases of this kind have a defined lifetime: they let developers try an interface before it settles, and they are withdrawn once a successor is in place.

Where it sits

Mariner is best read as Google's public answer to a question the industry was working through: how does a language model stop being a text box and start operating software? Browser use is the narrowest useful version of that problem — the browser is where most knowledge work already happens, and screen understanding is the most general way in. The prototype's benchmark score, its safety rules of thumb (active tab only, confirm before money moves) and its path into the Gemini API together describe the shift from an idea to a usable and governable category of product.