Google I/O: LLM capabilities power agentic AI search

Google has taken steps to advance artificial intelligence (AI) language models closer to what it calls “world models”, as it tries to make them more useful and universal.

The company used its annual developer event, Google I/O, to showcase the Gemini 2.5 large language model (LLM), new application programming interfaces (APIs) and programming tools and agentic AI functionality built into Google’s internet search engine. 

Gemini is Google’s primary AI engine, but it offers several others including Gemma 3n, a small language model for mobile devices.

Demis Hassabis, CEO of Google Deepmind, said: “Our ultimate vision is to transform the Gemini app into a universal AI assistant that will perform everyday tasks for us, take care of our mundane admin and surface delightful new recommendations – making us more productive and enriching our lives.”

Hassabis said the company was beginning to develop new AI capabilities, following on from work on a research prototype called Project Astra, which explored concepts such as video understanding, screen sharing and memory. “Over the past year, we’ve been integrating capabilities like these into Gemini Live for more people to experience today.”

Google has been working to make its main AI model, Gemini, a world model. With Gemini 2.5 Pro, Hassabis said the model can make plans and imagine new experiences by understanding and simulating aspects of the world.

Hassabis said the progress the company has made is based on training AI agents to master complex games such as Go and StarCraft, with its Genie 2 software able to generate 3D-simulated interactive worlds.

According to Hassabis, Gemini is making use of this work in how it handles world knowledge and reasoning to represent and simulate natural environments. Other examples include Veo, Google’s AI-based video content generator, which Hassabis said has a deep understanding of “intuitive physics”.

As it strives to make its AI more useful, the company has released a Gemini 2.5-powered feature called AI Mode, on its North American internet search site, to provide more in-depth querying than just what is possible with the AI Overview functionality currently available.

An agentic AI feature called Project Mariner is also now part of AI Mode, which Google said can help people searching the internet get tasks done quicker. As an example, Google said a query to find affordable tickets would use AI Mode to look across multiple websites, analysing hundreds of potential ticket options with real-time pricing and inventory, and handle the work of filling in forms.

“AI Mode will present ticket options that meet your exact criteria, and you can complete the purchase on whichever site you prefer, saving you time while keeping you in control,” Google said.

Another agentic AI feature uses AI Mode to fast-track browsing and purchases on websites, with the entire payment process automated using Google Pay.

To support software developers, Google has integrated Gemini 2.5 Pro into the native code editor of Google AI Studio, which it said would help programmers prototype faster.

It has also released a beta version of Jules, an asynchronous code agent, which works directly with a software developer’s GitHub repositories.

Google said users can ask Jules to take on tasks such as version upgrades, writing tests, updating features and bug fixes.