Amazon releases AI agents it says can work for days at a time

Included is a new category of AI agents that can carry out tasks for hours or days without getting stuck and asking users for help, AWS said.

“This is what we’ve been hard at work on over the last year: how do you build this really robust brain that can do complicated work streams," AWS Chief Executive Matt Garman said in an interview.

Agents are AI tools that work by taking some action on behalf of users. But in their current form, they often hit a wall and constantly need to come back to users for guidance.

Garman said the new “frontier agents," capable of performing tasks with broader goals for longer periods of time, are the product of what he called “an enormous amount" of software engineering and infrastructure data, a combination of different models and a strong underlying memory architecture.

AWS also shared news on a service called Nova Forge, which lets enterprises train private instances of Amazon’s Nova models on their own proprietary data, and the general availability of its Trainium3 AI chip.

AWS made the announcements at the company’s annual re:Invent conference in Las Vegas. Together, they highlight the cloud giant’s AI commitments amid concerns over the pace of enterprise AI adoption and the ability of hyperscalers to meet growing demand.

AWS has sometimes been criticized as a laggard in the AI space since it was slower to release its own models. And while Garman agrees that AWS “took half a step back" at the beginning of the AI boom to focus on building a broad and scalable platform for enterprises, he said the agentic trend has put the cloud giant right back front and center. That is because agents require access to business data, core applications and other systems, much of which is supported by AWS architecture, he said.

In its third-quarter earnings call in October, Amazon said revenue from AWS, the source of much of the company’s profits, climbed 20% in the three-month period—the fastest rate of growth since 2022.

Jason Andersen, vice president and principal analyst at Moor Insights and Strategy, agreed that AWS’s pervasiveness inside the enterprise does give its AI solutions an advantage over competitors. At the same time, there is such a growing “universe of options" that will make the AI tooling market increasingly competitive, he said.

Nova Forge, another AI service announced during the conference, addresses enterprise use of the company’s Nova generative AI models. Typically, enterprises fine-tune existing pretrained models with their own data to give it context on their company and business. But this strategy only takes them so far, Garman said.

“If you over train on post training, a lot of the models will forget some of the core reasoning and other pieces that they know," he said.

The new service lets customers mix their enterprise data into initial pretraining and mid-training runs of AWS’s Nova models, yielding a custom-made model that is trained on their own data and available exclusively to them if they choose. Nova Forge also includes new methods of post-training.

Garman said beta customers using these custom models found 40% to 60% improvements compared with other strategies, such as fine-tuning and retrieval augmented generation.

“We think a lot of enterprises and startups are going to want to use this as a core where they want the model to really understand their business and their workflows," he said.

One challenge for AWS, as well as its competitors, will be building enough infrastructure to deliver the AI services.

“Demand keeps skyrocketing," Garman said. “We’ve added 3.8 gigawatts of new data center capacity in the last 12 months alone. And we’re not slowing anything down. We’re only speeding that up."

Write to Isabelle Bousquette at isabelle.bousquette@wsj.com