AWS announced significant enhancements to its AgentCore AI agent building platform during the AWS re:Invent 2025 conference. This year’s theme is AI for the enterprise. One of its goals is to democratize AI to customers by providing them more control to customize AI agents tailored to their needs. After the high-octane keynote by AWS CEO Matt Garman, the announcements began. He focused on how AI agents may realize the “true value” of AI.
The conference showcased a range of new features and services designed to improve AI capabilities for businesses across various sectors. One of the marquee debuts was the Policy feature in AgentCore. This effective AI Governance tool allows developers to impose restrictions and guidelines for their AI agents. We hope this new feature will help increase user trust and deliver a key pillar of more responsible and effective governance over all AI interactions.
New Features in AgentCore
These improvements to AgentCore are the result of increased appetite and necessity for more sophisticated AI solutions, and we’re excited to lead the way. One of the coolest things about AgentCore is the Policy function. It provides developers the tools to establish easy-to-understand operational boundaries for AI agents, so these new digital assistants don’t go off the rails. We believe this capability will make AI interactions in enterprise settings much more safe and reliable.
AWS alleges that its AI agents can learn from users’ interactions and function autonomously for long stretches of time. This turning point represents a shift beyond legacy AI assistants. Instead, we have increasingly autonomous AI agents that can execute complex tasks on their own, often without oversight or guidance.
“AI assistants are starting to give way to AI agents that can perform tasks and automate on your behalf,” – Matt Garman, AWS CEO
Innovations in AI Training Technology
AWS recently launched a new generation of its AI training silicon, Trainium3, with stunning performance improvements. According to AWS, Trainium3 offers performance increases of four times or better for AI training. It increases inference speed while being more efficient than previous models. This huge increase in performance is accompanied by an equally large drop in operating energy use. Trainium3 reduces energy consumption by a whopping 40%. These kinds of advancements are likely to help enterprises save time and resources when it comes to training AI in a more sustainable way.
Alongside Trainium3, AWS launched UltraServer, an AI brain designed to maximize the potential of the new chip. This system is engineered to optimize performance for various applications, providing customers with high-speed computing power essential for advanced AI tasks.
Introducing Nova Forge
Another key announcement from AWS re:Invent 2025 was the launch of Nova Forge, a service that provides cloud customers with access to pre-trained, mid-trained, or post-trained AI models. This service allows businesses to train models using their proprietary data, thereby tailoring AI solutions to meet specific operational needs.
This week, AWS announced the Nova Forge model building platform to ease the model training process. This tool will help more companies focus their efforts to get the most out of AI. Our customers are able to hit the ground running with models that have already been created, all while having the flexibility to edit those models to meet their specific needs.
“This is where we’re starting to see material business returns from your AI investments,” – Matt Garman, AWS CEO
To make it easier for customers to evaluate their AI agents, AWS is providing 13 out-of-the-box evaluation frameworks. We believe these systems are valuable resources for public sector organizations to use when evaluating the performance of their AI solutions. They make certain these solutions fit within business goals.
