Top announcements from the first day of AWS re:Invent 2025: AI agents, Trainium3 and new Nova models
By ForGeeks
The annual Amazon Web Services technology conference — AWS re:Invent — has completed its first official day and has already delivered a ton of news.The main theme is predictable — artificial intelligence for business. But this year, AWS is focusing on tools that give companies more options for customizing their own AI agents — including one that AWS claims can learn from the user and operate autonomously for days.
And it’s the first time that AWS re:Invent — has announced a number of new tools that will be available to help companies customize their own AI agents.
AWS re:Invent 2025, which runs through Dec. 5, opened with a keynote from AWS CEO Matt Garman. He emphasized that it is AI agents that unlock the «real value» of artificial intelligence.
«AI assistants are giving way to AI agents that can perform tasks and automate processes for you» — he noted. — «This is where we start to see real business results from investments in AI». The key announcements from the first day are compiled below.
AWS unveiled the next-generation Trainium chip — Trainium3 — and the UltraServer AI system built on it. Briefly, the updated chip offers a four-fold performance boost in training and inferencing AI models, and power consumption has been reduced by 40 percent.The company also hinted that it is already working on Trainium4, which will be able to interoperate with Nvidia solutions.
AAWS also announced that it is working on a Trainium3 chip that will be able to interface with Nvidia solutions.
AWS has updated its AgentCore AI agent creation platform. One of the key new features — Policy in AgentCore, allowing developers to more easily set rules and constraints for agents.In addition, agents will now be able to log actions and remember user information. AWS also unveiled 13 out-of-the-box systems for evaluating AI agent performance.
The company announced three new «Frontier» Frontier AI agents. One —«Kiro autonomous agent» — can write code and adapts to a team’s style to then work almost entirely on its own for hours or even days.A second agent is designed for security tasks, including code reviews, and a third — for DevOps processes, such as preventing incidents when rolling out updates. Preview versions are available now.
AWS is expanding its Nova family of models with four new AI models: three text-based models and one multimodal model capable of generating text and images.The company also introduced the Nova Forge service — a platform for accessing models at different stages of training that AWS customers can pre-train on their own data. The key focus here — is on flexibility and personalization.
One of its partner companies, Lyft, shared the results of implementing Anthropic Claude models through Amazon Bedrock. The service uses an AI agent to handle driver and passenger inquiries.Lyft says average time to resolve issues has decreased by 87%, and driver use of the AI agent has increased by 70% year over year.
Lyft says it’s seen an 87% reduction in average time to resolve issues, and driver use of the AI agent has increased by 70% year over year.
Amazon also unveiled «AI Factories» — an infrastructure solution that allows large companies and government agencies to run AWS AI systems right in their own data centers.The suite is co-built with Nvidia and can use both Nvidia gas pedals and the company’s new proprietary Trainium3. This solution is focused on the challenges of data sovereignty — where it is important for organizations to have full control over information and not hand it over to external cloud services.
