Google Explores AI Computer Use With Project Mariner

Key Takeaways

• Google released Gemini 2.0 on Dec. 11.
• Built on the new AI model, Project Mariner lets Gemini interact with web browsers.
• Microsoft and Anthropic have introduced similar features enabling AI computer use.

When Google unveiled Gemini 2.0 on Wednesday, Dec.11, CEO Sundar Pichai described it as the firm’s “most capable model yet” that will help usher in a new “agentic era” for AI.

Joining other AI labs that are shifting the emphasis from chatbots to sophisticated AI agents that can use digital tools and carry out requests, Project Mariner is equipping Gemini 2.0 with the ability to use computers.

Beyond Chatbots

Project Mariner rides a wave of interest in AI assistants that can do more than just answer questions.

Microsoft made a significant move in this direction when it introduced new Copilot capabilities earlier this year. The firm boasted that the new independent agents “can automate and orchestrate complex, long-running business processes with more autonomy and less human intervention.”

A week later, Anthropic unveiled a Claude update that lets the AI control a computer’s cursor and input information via a virtual keyboard.

In a similar vein, Project Mariner “can understand and reason across everything on your browser screen,” letting it navigate complex websites in real time, Google said.

“The Future of Human-Agent Interaction”

The development of new, more capable agents that can interact with a potentially limitless array of digital interfaces promises to transform the way people use AI.

The field has advanced significantly in a short space of time.

When Claude unveiled its computer use demo in October, it scored 56% on WebVoyager, an AI benchmark that assesses agents’ ability to manipulate websites. Less than two months later, Project Voyager has pushed the bar much higher, scoring 83.5%. Kura’s multi-agent setup performs even better.

Google emphasized that Project Mariner is an early research prototype. But as the technology progresses, such initiatives will “explore the future of human-agent interaction.”

Moreover, web browsers are just the start. As action bots become more capable, they will be able to use a greater range of software applications. Meanwhile, the work of integrating and connecting different systems will become much easier.

User Adoption

While Project Mariner is currently only available to early testers, the race to deliver agents that can perform more complex tasks to a wider audience has already started.

Commenting on the new technology to CCN, SOCi’s Director of Market Insights Damian Rollison said the potential applications span a diverse range of use cases:

“Ideally, you’d want an AI agent to be able to book your flight, hotel, and rental car for you, not just look up travel info, or compile relevant statistics into a report or slide deck to save you the trouble.”

However, “it’s notable that Mariner is not yet publicly available and that Google acknowledges the mistakes it can still make,” he said and added:

“The state of the AI arms race being what it is, the company probably feels it must rush these products out in order to stay competitive, but their utility will only be proven by broad consumer adoption.”