Google Quietly Kills Project Mariner, Folding Its Browser AI Into Gemini Agent
Products·2 min read·Android Authority

Google Quietly Kills Project Mariner, Folding Its Browser AI Into Gemini Agent

Google shut down Project Mariner on May 4 without a public announcement, ending its 17-month run as a standalone browser-controlling AI and folding the technology into the Gemini API and the new Gemini Agent.

Share:

Google has discontinued Project Mariner, the experimental Chrome-based AI agent it unveiled at I/O 2025, with the project's landing page now carrying a quiet farewell message. The shutdown took effect on May 4, 2026 without a public announcement, ending a 17-month run during which Mariner served as Google DeepMind's flagship demo of a model that could browse the web like a human — clicking buttons, filling forms, and booking travel from sites like Expedia.

Mariner worked by taking continuous screenshots of the user's browser, recognizing buttons and text with a vision-language model, and then issuing clicks and keystrokes to drive the page. The approach was always compute-heavy: visual processing at full browser resolution required significant inference cost per action and was prone to errors such as selecting the wrong option on dropdown menus or misreading custom UI elements. Privacy concerns also dogged the project, since the agent needed continuous access to whatever was visible in the user's browser at any moment.

The signs were on the wall well before this week. Google reassigned the Mariner team in March 2026, redirecting headcount toward Gemini's core agentic capabilities. The decision lands in a market where the agent narrative has shifted decisively toward "digital coworker" tools — Anthropic's Claude Code and the open-source OpenClaw — that operate at the file and terminal layer rather than the browser pixel layer. Those tools sidestep the screenshot-and-click bottleneck that capped Mariner's reliability.

Google says Mariner's research will live on. Most of its capabilities are being merged into the Gemini API and a new product called Gemini Agent, which Google has positioned as a successor that takes higher-level instructions and routes them through the most appropriate tool surface — sometimes a browser, sometimes a structured API, sometimes a code execution environment. In practice, that means the visual web-control technology becomes one capability among many rather than a standalone product line.

The Mariner sunset also fits a broader pattern at Google through 2026 of consolidating experimental AI surfaces into a smaller number of flagship products centered on Gemini. The trade-off is one Google has made before: less brand sprawl and clearer messaging at the cost of losing distinctive product identities that researchers and early adopters had come to recognize. Whether Gemini Agent fills the gap will depend on how quickly it matches the agentic coding tools that have outpaced browser-control approaches over the last six months.

Related Articles