TORONTO — Google has announced a new version of Gemini, its flagship AI model, which it claims can better power “agents” to carry out tasks in the real and digital worlds. The move adds the search giant to a growing list of companies betting that such agents can become a major, practical application of generative AI.

In a blog post, Google CEO Sundar Pichai claimed the updated system is particularly well-suited for building agents, because it can respond faster, take in more complex instructions, make plans and use external tools.

Google is also opening up two assistants of its own for testing. The first, dubbed Project Mariner, operates within a browser window, taking in information that’s visible or hidden in a web page’s code. The agent can then carry out tasks that require typing and clicking, like filling a shopping cart with items. 

“AI assistants of the future will need to be able to navigate the web and take action on your behalf to be seen as truly helpful,” said Jaclyn Konzelmann, director of product management at Google Labs.

The other new agent, Project Astra, was first teased in May. The virtual sidekick can keep up a constant dialogue with a user, answering questions or providing location and other factual information from Google apps. “It engages in free-flowing conversation based on what you see and what you say, and responds at the speed of human conversation,” said Bibo Xu, a group product manager at Google’s DeepMind division.

Astra will initially hit the real world embedded in smart glasses. In a demonstration video shown to reporters, someone wearing the frames in London instructed the assistant to pull a door code from an email, then remember it for later use. Google eventually plans to make the agent available via its AI app.

The firm describes both Mariner and Astra as research prototypes, and did not provide timelines for when it will make them widely available.

Google is the latest builder of large language or foundation AI models to expand into the assistant space. Amazon on Monday announced a new San Francisco-based AI lab led by David Luan, who was hired when Adept, his agent startup, was acquired in a staff-and-licensing deal in June. 

Earlier this month, Toronto’s Cohere said it’s developing workplace AI helpers for businesses. In October, Anthropic—an Amazon-backed startup—launched a new AI model that can  control  computers to fill out forms among other actions. And OpenAI is reportedly planning to release an agent next year.

People may be more reluctant to delegate tasks to an AI assistant than to ask a chatbot questions. Google executives acknowledged concerns about the risks of agents, and promised the ones it’s rolling out won’t do anything significant without human say-so. For example, Mariner may be able to put things in an online store’s shopping cart, but it can’t complete the purchase.

Copyright holders are increasingly seeking to prevent AI developers from crawling their sites to train new models, and may have similar objections to agents. For now, Mariner can do most things a human could by clicking and typing, Konzelmann said. But Google plans to talk to publishers about “what controls really make sense in this space of having agents that can take actions on a user’s behalf on the internet.”