Google Unveils A.I. Agent That Can Use Websites on Its Own

Chatbots today can handle a variety of tasks, such as answering questions, composing poetry, and creating images. Looking ahead, they may gain the ability to autonomously complete tasks like online shopping or working with tools such as spreadsheets.

On Wednesday, Google introduced a prototype of this emerging technology, known among artificial intelligence researchers as an A.I. agent.https://wordpress.com/

Google is one of several tech companies developing A.I. agents. Startups like OpenAI and Anthropic have also released similar prototypes capable of interacting with software applications, websites, and other digital tools.

Google’s latest prototype, called Mariner, is built on Gemini 2.0, a system also unveiled on Wednesday. Gemini serves as the foundational technology for many of Google’s A.I. products and research efforts. Different iterations of Gemini power Google’s chatbot of the same name and A.I. Overviews, a tool in Google Search designed to provide direct answers to user queries.

“We’re enabling users to input requests into their web browsers and have Mariner carry out actions on their behalf,” said Jaclyn Konzelmann, a Google project manager, in an interview with The New York Times.

Gemini itself is a type of neural network — a sophisticated mathematical system that learns skills by processing large datasets. For example, by identifying patterns in vast collections of articles and books sourced from the internet, a neural network can develop the capability to generate coherent text.

The newest version of Gemini learns from diverse data sources, including text, images, and sounds. This training encompasses examples such as how people interact with spreadsheets, shopping platforms, and other online services. Using this knowledge, Mariner can perform similar tasks on behalf of users.

“It understands the need to press a button to trigger an action,” said Demis Hassabis, head of Google’s core A.I. lab, in the same interview. “It can operate in the real world.”

However, Mariner is designed to function with human oversight. For example, while it can fill a virtual shopping cart with groceries if a user has an active browser tab, it won’t complete the purchase. The user must finalize the transaction.

Sundar Pichai, Google’s CEO, stated in a blog post that these developments “bring us closer to realizing the vision of a universal assistant.”

Mariner has been designed as a Chrome extension, highlighting its strategic role in Google’s broader A.I. ambitions. However, this plan faces potential hurdles. Following a landmark ruling that deemed Google’s search engine an illegal monopoly, the U.S. Justice Department has requested a federal judge to mandate the sale or spinoff of Chrome.

There are other obstacles, too. Konzelmann acknowledged that Mariner, like other chatbots, occasionally makes errors. These systems rely on patterns within large datasets, which can sometimes lead to mistakes. While errors in text generation may go unnoticed, inaccuracies in executing actions online can pose significant challenges.

“Is it always accurate? Not yet,” Konzelmann admitted. “This is still experimental technology.”

Google is currently sharing Mariner with a limited group of external testers, with no timeline yet for a broader release.

On the same day, Google also unveiled an upgraded version of Project Astra, a smartphone digital assistant that responds to images, text, and voice commands. Similar to advancements introduced by OpenAI earlier this year, Astra represents a more sophisticated evolution of assistants like Apple’s Siri.

For now, Astra remains unavailable to the general public.