Google announced on Wednesday that Project Mariner will initially be released to a small group of preselected testers. This experimental AI agent is a major departure from paradigms about user experience (UX) because generative AI may interact with websites on behalf of users. Users can now command the AI to do specific tasks instead of having to navigate websites manually—a move that could reverberate far and wide in the business and publishing worlds of the web.

This work was demonstrated by Google Labs director Jaclyn Konzelmann. After installing an optional Chrome extension, users can open up a chat window in Chrome, where they can give such commands as, “Create a shopping cart from a grocery store based on this list.” Once this is done, the browser will automatically visit the related web site and search for the specified goods or services, adding them into a virtualcart.

However, the system is not without its limitations. The agent’s response time is somewhat slow, with noticeable delays of around five seconds between cursor movements. Furthermore, when details are ambiguous, like the quantity of items for carrots, Project Mariner pauses to clarify.

How does Project Mariner work?

Behind the scenes, Project Mariner relies on a combination of browser screenshots and cloud-based processing. The AI takes snapshots of the user’s browser, sends them to Gemini for analysis, and then receives instructions to execute actions on the website. Importantly, users must agree to this functionality via Google’s terms of service.

Learn more about Gemini Here

While the AI can search and interact with websites, it avoids some activities, such as filling out payment details, accepting cookies, or signing terms of service agreements. Google explains that these limitations are necessary to ensure users have control over sensitive activities.

For now, Project Mariner only works within the context of an active Chrome tab. This means the user cannot multitask on the computer while the AI runs its course. Google DeepMind’s CTO said this is done in terms of transparency. The end users can monitor step-by-step what the AI agent is actually doing.

Impact of Project Mariner on the Web

Project Mariner raises a host of critical questions for publishers and commercial ventures reliant on direct user interaction. Even as the AI surfs through websites in real time—guaranteeing that web traffic is still delivered—the users may be spending less time actually interacting with the sites themselves. In the future, Google’s agents will likely skip right over these traditional interfaces.

Konzelmann noted that this is a key UX shift, pointing out the requirement to rethink how users, websites, and AI agents interact: “We need to figure out what is the right way for all of this to change the way users interact with the web, and the way publishers create experiences for users, as well as for agents, in the future.”

Besides Project Mariner, Google announced other AI-specific agents designed for specific usage cases.

Deep Research is designed to assist users in exploring complex topics by breaking them into actionable, multistep plans. This agent competes with OpenAI’s offerings but focuses primarily on information gathering. When prompted with a large question, Deep Research generates a detailed plan, searches the web, and compiles a comprehensive report. It is currently available in Gemini Advanced and will roll out to the Gemini app in 2025.

Jules is another AI agent that targets developers by directly integrating into GitHub workflows. Jules can analyze existing code, suggest improvements, and implement changes. Currently in beta, Jules is expected to be widely available in 2025.

Google DeepMind is also working on AI agents that can be used to navigate virtual environments. Collaborating with developers such as Supercell, the company is testing Gemini to interpret and interact with the gaming worlds of Clash of Clans. This technology is still in prototype but will be useful for the future extension of AI capabilities to physical and virtual navigation.

Google has not released the public date of Project Mariner, but it holds some enormous implications from the two AI agents mentioned above. They challenge traditional human-centric web design towards a future where generative AI can do tasks for people. Companies, publishers, and developers have to gear up for the change, where AI agents begin to reframe web navigation, research, coding, and even games.