Google DeepMind's Project Mariner: AI Agents That Navigate the Web Like Humans
Google DeepMind unveiled Project Mariner, a research initiative developing AI agents capable of navigating websites, filling forms, making purchases, and completing multi-step web tasks autonomously using Gemini's vision capabilities.
Google DeepMind has publicly demonstrated Project Mariner, an AI agent system that uses Gemini's multimodal capabilities to navigate the web like a human user. The agents can understand web pages visually, click buttons, fill forms, navigate between sites, and complete complex multi-step workflows.
Project Mariner builds on Gemini 2.5's native vision and reasoning capabilities, combined with a specialized web interaction layer that translates agent intentions into browser actions.
Demonstrated Capabilities:
- •Booking flights by comparing prices across multiple airline websites
- •Filling out complex government forms by extracting information from uploaded documents
- •Managing email workflows: reading, categorizing, drafting responses, and scheduling follow-ups
- •Shopping across e-commerce sites with price comparison and coupon application
- •Research tasks spanning 20+ web sources with citation tracking
The system is currently in limited preview with select Google Workspace Enterprise customers. Google emphasized that all agent actions require user approval for sensitive operations (purchases, form submissions) and that agents operate in a sandboxed browser environment.
"Project Mariner represents a shift from AI that answers questions to AI that takes action," said the DeepMind team lead. "The web was built for humans to navigate — now we're teaching AI to navigate it too."
Analysts note that web-navigating agents could fundamentally change how businesses interact with online services, potentially disrupting the SaaS industry by enabling AI to use existing tools rather than requiring API integrations.