Google Project Mariner: AI Agents to Navigate the Web

Google Introduces Project Mariner: A New AI Agent for Web Interaction

On Wednesday, Google revealed Project Mariner, its inaugural AI agent capable of performing actions directly on the internet. This research prototype, originating from the company’s DeepMind division, represents a significant advancement in artificial intelligence.

How Project Mariner Functions

Powered by the Gemini model, the agent assumes control of a user’s Chrome browser. It simulates human interaction by manipulating the cursor, selecting buttons, and completing online forms, enabling it to utilize and traverse websites in a manner akin to a person.

Initially, Google is releasing this AI agent to a limited number of testers who were preselected. Access will be expanded gradually.

A Paradigm Shift in User Experience

Google continues to explore innovative methods for Gemini to interpret, summarize, and now actively utilize websites. A Google executive explained to TechCrunch that this represents a “fundamentally new UX paradigm shift.” The goal is to move users away from direct website interaction and towards engaging with a generative AI system that handles these tasks on their behalf.

This evolution has the potential to impact a vast number of businesses, including publishers such as TechCrunch and retailers like Walmart, who traditionally depend on Google to drive traffic to their online platforms.

Demonstration and Capabilities

During a demonstration for TechCrunch, Jaclyn Konzelmann, director of Google Labs, showcased Project Mariner’s functionality.

After installation as a Chrome extension, a chat interface appears alongside the browser. Users can then instruct the agent to perform tasks, for example, “create a shopping cart from a grocery store based on this list.”

The AI agent then navigates to the specified website—in this instance, Safeway—searches for items, and adds them to a virtual cart. It’s noteworthy that the agent operates at a relatively slow pace, with approximately 5 seconds elapsing between each cursor movement.

Limitations and Security Measures

The agent is currently unable to complete checkout processes, as it is designed to avoid handling sensitive information like credit card details or billing addresses. Furthermore, Project Mariner will not accept cookies or agree to terms of service agreements.

Google has intentionally implemented these restrictions to maintain user control and ensure data security.

Technical Implementation

The agent functions by capturing screenshots of the browser window—a process requiring user consent outlined in the terms of service—and transmitting them to Gemini in the cloud for analysis. Gemini then sends instructions back to the computer to control web navigation.

Project Mariner can also assist with tasks like finding flights and hotels, shopping for household goods, and discovering recipes, all of which typically require manual web browsing.

Operational Constraints

A key limitation is that Project Mariner operates exclusively within the currently active tab of the Chrome browser. This means users cannot simultaneously perform other tasks on their computer while the agent is working; continuous observation of Gemini’s actions is necessary.

Koray Kavukcuoglu, chief technology officer at Google DeepMind, emphasized that this design choice was deliberate, ensuring users remain aware of the AI agent’s activities.

Implications for Website Owners

Website owners may find reassurance in the fact that Google’s AI agent operates on the user’s computer screen, meaning publishers and retailers still receive direct views of their pages. However, the agent could potentially lead to reduced user engagement with websites, and ultimately, may diminish the need for direct website visits altogether.

Future of Web Interaction

“This [Project Mariner] represents a fundamentally new UX paradigm shift that we are witnessing,” Konzelmann stated to TechCrunch. “We must determine the optimal approach for this to reshape how users interact with the web, and how publishers can create experiences for both users and agents in the future.”

Additional AI Agents Unveiled by Google

Alongside Project Mariner, Google also introduced several other specialized AI agents on Wednesday.

Deep Research: For Complex Topic Exploration

Deep Research is designed to assist users in exploring intricate subjects by formulating multistep research plans. It shares similarities with OpenAI’s o1, which also offers multistep reasoning capabilities.

However, a Google spokesperson clarified that this agent is not intended for solving mathematical or logical problems, writing code, or performing data analysis. Deep Research is currently available in Gemini Advanced and will be integrated into the Gemini app in 2025.

When presented with a challenging or extensive question, Deep Research generates a detailed action plan for answering it. Upon user approval, the agent undertakes a few minutes of research and web searching, culminating in a comprehensive report.

Jules: An AI Coding Assistant

Jules is a new AI agent aimed at aiding developers with coding tasks. It seamlessly integrates with GitHub workflows, allowing Jules to access existing code and make direct modifications within the GitHub environment.

Jules is currently being tested by a select group of beta users and will become more widely available later in 2025.

Game-Playing AI Development

Google DeepMind is also developing an AI agent to assist players in navigating video games, leveraging its extensive experience in creating game-playing AI. The company is collaborating with game developers, such as Supercell, to evaluate Gemini’s ability to interpret gaming environments like Clash of Clans.

While a release date for this prototype remains unspecified, Google believes this work will contribute to the development of AI agents capable of navigating both virtual and physical worlds.

The Future Impact of AI Agents

The widespread rollout of Project Mariner to Google’s extensive user base remains uncertain. However, when it occurs, these agents are poised to have a substantial impact on the broader web landscape.

The internet was originally designed for human interaction, but Google’s AI agents have the potential to fundamentally alter this paradigm.