COSMO Browser Agent &
Project Mariner Explained

How Google DeepMind's browser automation initiative may power one of COSMO's most ambitious capabilities — autonomous web interaction.

📅 Last updated: May 2, 2026
ⓘ The connection between COSMO and Project Mariner is speculative — not officially confirmed

What Is Project Mariner?

Project Mariner is an initiative from Google DeepMind focused on building AI agents that can autonomously navigate and interact with websites. Announced in late 2024, Mariner demonstrated the ability to understand web page content, click buttons, fill forms, and complete multi-step tasks across different websites — all without human intervention.

Unlike simple web scraping or macro tools, Mariner uses AI to understand web pages contextually — interpreting layouts, reading content, and making decisions about how to accomplish user-defined goals.

The COSMO Connection

COSMO's reported Browser Agent skill closely mirrors the capabilities demonstrated by Project Mariner. While Google has not officially confirmed the connection, the technology stack aligns:

  • Autonomous browsing: Both Mariner and COSMO's Browser Agent reportedly navigate websites independently
  • Multi-step tasks: Both handle complex workflows spanning multiple pages and interactions
  • Contextual understanding: Both interpret page content rather than relying on predefined selectors
  • Google DeepMind origin: COSMO's package name includes "research," suggesting a research-origin project

Potential Use Cases

🛒

Shopping & Comparison

Autonomously browse multiple retailers, compare prices, and find the best deals based on user preferences.

📋

Form Filling

Complete online forms, applications, and registrations using stored user information and context.

🔄

Workflow Automation

Execute complex multi-step tasks like booking appointments, managing subscriptions, or processing returns.

📊

Data Gathering

Collect and organize information from multiple websites into structured summaries and reports.

How Browser Agent Might Work

Based on Project Mariner demonstrations and COSMO's reported architecture, the Browser Agent likely operates through a pipeline:

  1. User intent: The user describes a goal (e.g., "find the cheapest flight to Tokyo next month")
  2. Task planning: COSMO decomposes the goal into discrete steps
  3. Web navigation: The Browser Agent opens relevant sites and navigates through search, filters, and results
  4. Data extraction: Results are collected, compared, and structured
  5. User presentation: A summary with actionable options is presented to the user

Privacy and Safety Considerations

Autonomous web browsing raises significant privacy and safety questions. If COSMO's Browser Agent can interact with authenticated websites, it may have access to sensitive accounts and personal data. Key considerations include:

  • How are login credentials handled? Does COSMO store or access passwords?
  • Can users set boundaries on which sites the agent can access?
  • How are financial transactions handled — can the agent make purchases?
  • What guardrails prevent unintended actions on sensitive websites?

For more on COSMO's data handling, see our Privacy & Permissions guide.

All Features →COSMO Skills →COSMO vs Gemini →

📚 Sources

  • Google DeepMind — Project Mariner official page and capabilities
  • Google Labs — Project Mariner Help documentation
  • 9to5Google — COSMO Browser Agent skill reporting from Play Store analysis