Google Started Project Mariner for Multi-Tasking AI Agents

Imagine telling an AI to search for flights, check three hotel sites, and fill out a booking form while you work on something else. That was the promise of Google Project Mariner, an experimental agent from Google DeepMind that could browse the web and take actions for you. It also became one of the first mainstream examples of a multi-tasking AI agent, one that could handle several jobs at the same time. But the story did not end the way many people expected. Google shut down the standalone product in May 2026 and moved its technology into other tools.

This guide explains what Project Mariner was, how it worked, why it matters, and where its ideas live now. If you want to understand the bigger shift toward AI systems that plan and act on their own, a program like the Certified Agentic AI Expert course covers how these agents are designed and used. The language below is simple enough for beginners and detailed enough for professionals.

Quick Answer: What Was Project Mariner?

Project Mariner was an experimental AI agent built by Google DeepMind to use a web browser the way a person does. It could look at a page, click buttons, type into forms, scroll, and complete multi-step tasks such as searching listings or booking travel. In May 2025, Google expanded it so that it could run up to 10 tasks at once on cloud-based virtual machines. Google quietly shut down Project Mariner as a standalone product on May 4, 2026, and its core features are now part of Gemini Agent, Chrome's auto browse feature, the Gemini API, and AI Mode.

How Project Mariner Worked

The Basic Idea

Most chatbots only talk. Project Mariner could act. You gave it a goal in plain language, such as "find three job listings that match my resume," and it opened web pages, read them, and took steps toward the goal. You could watch what it was doing and step in at any time.

How It "Saw" the Web

Reports describe Mariner as screenshot-based. It captured pictures of the browser window, used visual recognition to understand buttons, text, and forms, and then clicked or typed accordingly. It was built on Google's Gemini models, starting with Gemini 2.0 and later using Gemini 2.5 capabilities.

The Data and Model Side

Under the hood, agents like this depend on trained models, testing data, and careful evaluation to avoid mistakes. Professionals who build and test such systems often have strong data skills, and programs such as the Certified Data Science Developer course cover the data handling, modeling, and analysis that sit behind AI products.

Cloud Virtual Machines

In the 2025 version, tasks ran on cloud virtual machines instead of only inside your own browser tab. This freed you to keep browsing or working while the agent handled tasks in the background.

Project Mariner Timeline

What Multi-Tasking Meant for AI Agents

Doing Several Jobs at Once

The big upgrade in 2025 was parallel work. Instead of finishing one task before starting another, Mariner could handle up to 10 tasks at the same time. For example, you could ask it to compare prices on several products, book a table, and gather research, all in one go.

Teach and Repeat

Google also described a feature that let you show the agent a workflow once so it could repeat similar tasks in the future. This idea, learning from a demonstration, is a core goal of modern agent research.

Why This Matters

Multi-tasking turns an assistant from a single helper into something closer to a small team. It also raises new questions about supervision, accuracy, and safety, because mistakes can multiply when many tasks run at once.

Who Could Use It?

Access was limited. Mariner was available to Google AI Ultra subscribers in the United States, a plan reported at $249.99 a month, plus developers through the Gemini API and Vertex AI. That kept it out of reach for most consumers and kept its audience small.

Why Did Google Shut Down Project Mariner?

Google has not published a long explanation. It told users that the tool's technology "voyaged to other Google products" and pointed people to Gemini Agent for complex tasks. Reports and analysts suggest several reasons, though these are interpretations, not confirmed statements.

Consolidation

Google appears to be bringing its agent efforts under the Gemini umbrella. Instead of several separate experiments, it wants one agent system across the Gemini app, Chrome, Search, and developer tools. Reports also say Mariner's team moved to a new group building a more advanced agent.

Limits of Screenshot-Based Agents

Agents that read screens can be slow and fragile. A small change to a website can confuse them. Many in the industry have moved toward agents that use direct tools, APIs, and code, which can be faster and more reliable than clicking through web pages.

Shifting Industry Focus

Several reports noted that attention has moved toward coding and general-purpose agent tools. Competing browser agents also exist from other companies, so Google has been merging its own capabilities into products that more people already use.

Cost and Reach

A premium plan with limited availability made it hard for Mariner to reach large audiences, while integrating the technology into mainstream products can serve far more users.

Where Mariner's Technology Lives Now

Gemini Agent

Gemini Agent, available in the Gemini app for paid subscribers, is described as the direct successor. It can handle tasks such as archiving emails, making reservations, and working through multi-step workflows, and Google has said it builds on insights from Mariner.

Chrome Auto Browse

In January 2026, Google added a Gemini side panel to Chrome with an auto browse feature that lets an AI agent act inside the browser. Google has not said exactly how much of this came from Mariner, but analysts see it as part of the same effort.

Gemini API and Computer Use

Developers can access computer use capabilities through the Gemini API, which lets them build their own agents that operate software interfaces.

AI Mode in Search

Some agentic features, such as helping with tasks like reservations, have been brought into AI Mode in Google Search.

Gemini Enterprise Agent Platform

In April 2026, Google announced the Gemini Enterprise Agent Platform as the evolution of Vertex AI, with tools for building, deploying, and governing agents in production. Reports from Google I/O 2026 also point to enterprise controls such as role-based access and audit logging.

Browser Agents vs Other Kinds of AI Agents

The trend is toward combining these methods, using direct tools when possible and falling back to screen control when no other option exists.

Risks and Safety Concerns

Agents that act on your behalf carry real risks.

• Mistakes at scale: A wrong click can book the wrong trip, and running many tasks multiplies errors.

• Prompt injection: Malicious text hidden on a website can try to trick an agent into doing something harmful.

• Privacy: Agents may see sensitive pages, such as email or banking.

• Payments: Letting an agent buy things needs strict limits and confirmation steps.

• Accountability: It can be unclear who is responsible when an agent makes a costly error.

Good practice includes keeping a human in the loop, requiring approval for purchases and sensitive actions, limiting access to only what the agent needs, and reviewing logs.

What This Means for Businesses and Professionals

Where Agents Can Help

• Researching suppliers and comparing prices

• Filling repetitive forms and updating records

• Managing bookings and calendars

• Gathering and summarizing information from many sites

Skills That Will Matter

Understanding how agents plan, use tools, and fail will be a valuable skill in many jobs. Because AI agents sit alongside other advanced technologies such as automation, data systems, and secure platforms, many professionals build a broader profile. You can explore a wide range of options through Deep Tech Certification programs, which cover advanced and emerging technologies beyond a single product.

How to Prepare for the Age of AI Agents

• Learn the basics. Understand what agents are and how they differ from chatbots.

• Start small. Test agents on low-risk tasks first.

• Set clear rules. Decide what the agent may and may not do.

• Keep approvals in place. Require confirmation for purchases and data changes.

• Check outputs. Verify results, especially for money, legal, or health decisions.

• Protect your data. Use separate accounts and limited permissions where possible.

• Stay current. Products change quickly, and features move between tools, as Mariner showed.

Common Mistakes to Avoid

• Trusting an agent blindly. Always review important actions.

• Giving too much access. Limit permissions to what is needed.

• Assuming a product will last. Experimental tools can disappear, so avoid building critical workflows on them without a backup plan.

• Ignoring security. Agents that browse the web face unique attack risks.

• Expecting perfection. Even strong agents can misread pages or misunderstand goals.

• Skipping testing. Run trials before using an agent at scale.

Simple Tips for Beginners

• Think of an agent as a helpful assistant that still needs supervision.

• Write clear, specific instructions.

• Try tasks that are easy to check.

• Never share passwords or payment details unless you fully trust the setup.

• Keep notes on what works and what fails.

• Follow official announcements from Google and other providers.

• Learn the basics of AI so you can judge claims calmly.

Conclusion

Google Project Mariner was a bold early step toward multi-tasking AI agents. It showed that an AI could see a browser, take actions, and manage several jobs at once. Although Google closed the standalone product on May 4, 2026, the idea did not disappear. Its ideas now live in Gemini Agent, Chrome's auto browse, the Gemini API, AI Mode, and Google's enterprise agent tools. The lesson for anyone watching this space is simple: agent products will keep changing, so build understanding and good habits rather than depending on a single tool.

Great technology also needs clear communication. If you want to learn how to explain and promote AI products responsibly and reach the right audience, a Marketing Certification is a smart next step.

FAQs

1. What is Google Project Mariner?

Project Mariner is an experimental AI agent developed by Google that explores how AI can interact with websites and complete tasks on a user's behalf. It was introduced as a research prototype built with Gemini 2.0 and focused initially on browser-based computer use.

2. Why did Google start Project Mariner?

Google developed Project Mariner to explore a new generation of AI agents capable of understanding digital environments and taking actions rather than simply generating text. The project investigates how AI can interact with browsers, websites, forms, and other web elements to accomplish practical tasks.

3. What can Project Mariner do?

Project Mariner can understand information displayed in a browser and interact with web pages using actions such as clicking, scrolling, and typing. Google has demonstrated use cases including research, information lookup, bookings, and purchasing-related tasks.

4. How does Project Mariner support multitasking?

An updated version of Project Mariner introduced a system of agents capable of working on up to ten different tasks at the same time. This allows multiple activities to be handled concurrently rather than requiring one task to finish before another begins.

5. Is Project Mariner a multi-agent AI system?

The updated Project Mariner uses a system of agents designed to handle multiple tasks simultaneously. This represents a shift from a single-task assistant toward an agentic system capable of coordinating several independent activities.

6. How does Project Mariner interact with websites?

Project Mariner analyzes information from a browser, including visual elements and web content, and uses computer-use capabilities to interact with the active webpage. Its actions can include typing, scrolling, and clicking.

7. Is Project Mariner powered by Gemini?

Yes. Google initially described Project Mariner as an early research prototype built with Gemini 2.0. Its computer-use capabilities were later being brought to developers through the Gemini API.

8. What types of tasks can Project Mariner perform?

Project Mariner can assist with tasks such as researching information, navigating websites, making certain bookings, and shopping-related activities. The goal is to allow an AI agent to perform multi-step digital tasks rather than simply provide instructions to the user.

9. Can Project Mariner make purchases?

Project Mariner was designed to support tasks involving purchases, but Google introduced safeguards around sensitive actions. In its early prototype, the system required user confirmation before certain sensitive actions, including purchasing something.

10. What is the "teach and repeat" feature in Project Mariner?

"Teach and repeat" is a capability Google demonstrated that allows users to show an agent how to perform a task once. The system can then learn a plan that can be applied to similar tasks in the future.

11. Is Project Mariner available to everyone?

Availability has changed as Google has developed the technology. In its May 2025 update, Google said the updated Project Mariner was available to Google AI Ultra subscribers in the U.S., while its computer-use capabilities were also being brought to developers through the Gemini API.

12. What is computer use in Project Mariner?

Computer use refers to the ability of an AI system to interact with a graphical computer environment in ways similar to a human user. In Project Mariner, this includes interpreting browser content and performing actions such as clicking, scrolling, and typing.

13. How is Project Mariner different from a traditional chatbot?

A traditional chatbot primarily responds to questions or generates content. Project Mariner is designed to go further by interacting with digital environments and taking actions to complete multi-step tasks on behalf of the user.

14. How did Project Mariner perform in early testing?

In Google's initial announcement, Project Mariner achieved an 83.5% result on the WebVoyager benchmark in a single-agent setup. Google also emphasized that the system was still an early prototype and could be inaccurate or slow when completing some tasks.

15. What safety measures does Project Mariner use?

Google has emphasized keeping humans involved when agents perform potentially sensitive actions. In the early Project Mariner prototype, the agent was restricted to interacting with the active browser tab and required final confirmation for certain sensitive activities.

16. Can Project Mariner work on multiple websites?

Its browser-based computer-use approach is designed to navigate web pages and interact with web elements. The broader goal is to allow agents to perform tasks across the web, although capabilities and access can depend on the specific implementation and availability.

17. How does Project Mariner relate to Google's AI agents?

Project Mariner represents an early research step toward Google's broader agent ecosystem. Google has subsequently incorporated agentic capabilities into products such as Search and Gemini and has continued developing systems that can perform tasks on behalf of users.

18. Can developers build with Project Mariner technology?

Google announced plans to bring Project Mariner's computer-use capabilities to developers through the Gemini API. This allows developers to explore building applications and automation workflows around computer-using AI agents.

19. What are the potential benefits of multi-tasking AI agents?

Multi-tasking agents could reduce repetitive digital work by handling several activities concurrently. They may help with research, scheduling, shopping, data gathering, and other workflows that normally require users to navigate multiple websites manually.

20. What is the future of Project Mariner and multi-tasking AI agents?

Project Mariner helped demonstrate Google's direction toward AI systems that can understand digital environments, use tools, and perform actions. Google has continued expanding agentic capabilities across Search and Gemini, suggesting that computer-using and multi-tasking agents will remain an important area of AI development.