Project Mariner is a browser agent research prototype from Google DeepMind. Google introduced it on December 11, 2024 as part of the Gemini 2.0 launch and described it as an early experiment in human-agent interaction inside the browser.[1]
The system observes what is displayed in a browser, reasons about the user's goal, plans the required steps, and then interacts with websites to carry out the task. Google's current product page still describes it as a research prototype, even though it is already available in the United States to Google AI Ultra subscribers.[2][3]
| Date | Milestone |
|---|---|
| December 11, 2024 | Google introduced Project Mariner as a Gemini 2.0 research prototype with an experimental Chrome extension.[1] |
| May 20, 2025 | Google said the updated system could complete up to ten tasks at a time and made it available to Google AI Ultra subscribers in the U.S.[3] |
| 2025-2026 | Google DeepMind product pages said Mariner's computer-use capabilities were being brought into the Gemini API.[2] |
Google's product page breaks Project Mariner into three core stages: observing, planning, and acting.[2]
| Stage | Description |
|---|---|
| Observes | Identifies text, code, images, forms, and other browser elements |
| Plans | Interprets the goal and outlines the steps it will take |
| Acts | Navigates and interacts with websites while keeping the user informed |
The current page also emphasizes a "teach and repeat" workflow in which agents can try to reuse a previously learned browser routine with less user input.[2]
At launch, Google said Project Mariner achieved an 83.5% single-agent result on the WebVoyager benchmark. The same announcement also stressed that the system was limited to the active browser tab and asked for confirmation before sensitive actions such as purchases.[1]
Google kept those safety themes in later updates, describing Project Mariner as an experimental system that is still being improved through trusted-tester feedback and gradual rollout.[2][3]
As of the current Google DeepMind product page, Project Mariner is available in the U.S. to Google AI Ultra subscribers. Google also says it plans to bring the prototype's computer-use capabilities into the Gemini API and other Google products.[2][3]