Google is significantly advancing its AI agent capabilities, announcing a broader rollout and key updates to Project Mariner during its Google I/O 2025 conference. This experimental AI, designed to browse and interact with websites on a user’s behalf, now operates on cloud-based virtual machines and can manage up to ten tasks simultaneously. This development signals Google’s intensified push to reshape web interaction, empowering AI to actively perform tasks, a move that could fundamentally alter how users and businesses engage with the internet.
Initially, Project Mariner is available to U.S. subscribers of Google’s new $249.99-per-month AI Ultra plan, with support for more countries anticipated soon. Crucially for developers, Google is integrating Project Mariner’s functionalities into the Gemini API and Vertex AI, with broader availability expected this summer.
This integration will provide tools for creating more autonomous digital experiences, from purchasing tickets to ordering groceries without direct website visits. Testing partners like Automation Anywhere and UiPath are already building solutions with these capabilities.
The core of Project Mariner, as Google DeepMind describes, is its ability to observe what’s displayed in the browser, then reason to interpret goals, make a plan, and take action. This involves identifying and understanding diverse web elements like text, code, images, and forms to build a comprehensive understanding of on-screen content. A “teach and repeat” system further allows the AI to learn tasks after a single demonstration.
Enhanced Capabilities and User Experience
A key upgrade for Project Mariner is its transition to running on virtual machines in the cloud. This allows the AI agent to complete tasks in the background while users continue with other work—a notable improvement from its predecessor which ran directly on the user’s browser and limited concurrent activity. Google states the new iteration can handle up to ten tasks at once, a substantial boost in efficiency.
Users will eventually be able to access Project Mariner through AI Mode on Google Search, initially via Search Labs, with a demonstration showing AI Mode swiftly finding and reserving baseball game tickets. Google is collaborating with partners like Ticketmaster, StubHub, Resy, and Vagaro to facilitate some of these agent-driven interactions.
Alongside this, Google also demonstrated an early version of “Agent Mode” in the Gemini app. Google describes Agent Mode as a system where users can simply state their objective, and Gemini will then “intelligently orchestrate the steps to achieve it.”
The company further explains that Agent Mode integrates advanced features like live web browsing, research, and app integrations to manage complex tasks with minimal user oversight. (Google’s approach to such technologies is outlined in their AI Responsibility Principles.) This mode will soon help users with complex tasks like apartment hunting, including booking tours using tools like Zillow.
Navigating a Competitive AI Agent Landscape
Google’s Project Mariner enters an increasingly competitive field of AI-driven automation. OpenAI broadened access to its AI agent, Operator, for ChatGPT Pro users in February 2025, allowing automation of digital workflows, though Operator requires explicit user approval before executing actions. Microsoft also previewed a “computer use” feature within its Copilot Studio in April 2025, aimed at enterprise automation.
Anthropic has also been a key player, upgrading its Claude 3.5 Sonnet model in October 2024 that allows it to manage desktop tasks. These tools, including Amazon’s Nova Act, are largely in experimental stages, and early experiences have sometimes found them to be slow or prone to errors. Project Mariner’s integration into Gemini’s Agent Mode leverages advanced vision-language models to interpret visual elements, a key differentiator.
Performance Realities and Broader AI Strategy
Despite rapid advancements, the practical effectiveness of current AI agents in autonomously handling complex professional duties remains under scrutiny. A Carnegie Mellon University study offered a sobering assessment of AI agent capabilities.
The study found that even leading AI fully completed only 24% of assigned tasks. Researchers pointed to issues such as “a lack of common sense, poor social skills, and incompetence in web browsing,” concluding that while AI agents can assist, they are “likely not a replacement for all tasks at the moment,” according to the study.
Google’s own exploration into AI agents extends beyond Project Mariner. Reports from early May 2025 suggested Google was testing a “Computer Use” function within its AI Studio, potentially linked to its Gemma 3 models and streamlined Cloud Run deployment. This aligns with Google Cloud’s broader strategy, which has referred to multi-agent AI systems as the “next frontier.”
Google CEO Sundar Pichai emphasized this vision, describing agents as systems that “combine the intelligence of advanced AI models with access to tools, so they can take actions on your behalf and under your control,” at Google I/O. As Google continues to develop these agentic technologies, their success will hinge not only on technical prowess but also on addressing reliability, safety, and the nuanced challenges of real-world task automation.
Last Updated on May 26, 2025 1:14 pm CEST
