Google today unveiled Project Mariner, its first AI agent capable of autonomously navigating web browsers, operating through a Chrome extension that controls cursor movements and form-filling to replicate human interactions online.

The Gemini-powered prototype, developed by Google’s DeepMind division, is initially available to a select group of testers. During demos, the agent performed tasks like creating shopping carts on grocery websites, though with noticeable five-second delays between actions. The system captures browser screenshots and processes them through Gemini in the cloud to generate navigation commands.

It operates only in Chrome’s active tab, requiring users to observe its actions rather than running in the background. Project Mariner achieved an 83.5% success rate on the WebVoyager benchmark for web-based tasks. The agent has built-in limitations, including inability to complete purchases, accept cookies, or agree to terms of service. Google Labs Director Jaclyn Konzelmann described the project as a “fundamentally new UX paradigm shift” that could transform how users interact with websites. The company said it is engaging with web ecosystem stakeholders as development continues.