What is the difference between a generative AI model and an AI agent? The former generates text, code, or images based on prompts. The second is setting goals, planning, using tools, and reading and writing business applications and running and validating them all to achieve those goals. It is this gap between content generation and focused action that is at the heart of Amazon Web Services' strategy. No longer an introduction to templates, it's a platform and tools for enterprises to build customized agents that can be trained on data, managed deterministically, and operationalized quickly.
“With the advent of AI agents, we believe we've reached a tipping point in the trajectory of artificial intelligence from technological wonder to real value. Agents are exciting because they can take action and get things done,” CEO Matt Garman said in the opening keynote at re:Invent 2025, an annual event outlining strategy and news for Amazon's cloud division. This year, we are betting all our cards on artificial intelligence, or better yet, agent AI, which will allow us to build a complete and functional ecosystem. Enterprises can easily and quickly set up agent systems without having to deal with infrastructure or models. This is a ready-to-use SDK that means for AWS to engage and keep companies large and small in its ecosystem.
The foundation is the Bedrock AgentCore platform, which provides the “building blocks” for building capable and manageable agents. All cutting-edge GenAI models are available, from ChatGPT to group-invested Anthrtopic's Claude to new models from European Mistral, allowing users to focus on their applications and use cases. Two automated features are added at this stage: a policy feature that imposes limits on the agent's behavior, and an evaluation feature that monitors the agent's behavior in the real world for accuracy, usefulness, and security.
At the infrastructure level, this service is powered by Ai Factories, which respects sovereignty, policy, and governance and aims to deploy Aws services into customers' existing data centers using their own data assets. At the same time, Garman announced the launch of its new Trainium 3 chip, which is twice as energy efficient as version 3 for large-scale training and global inference. Officially, the product aims to integrate with the Nvidia family, but on a strategic level it appears to want to reduce its dependence on the AI chip giant. Aws' Agentica platform is based on the new Nova 2 family of models and powered by Omni, an integrated multimodal model that can integrate voice, text, video, and images. Building on these models is Nova Forge, a service that allows you to create customized Frontier models.
The culmination of this vision is precisely the Frontier Agent class. In other words, it is an autonomous, scalable, long-running agent that can work in conjunction with a team and coordinate disparate tools and data without ongoing intervention. This is where the “Agent SDK” becomes a competitive advantage. Policy, evaluation, and memory features provide reliability, while AgentCore ensures resiliency, the power and training capabilities of Trainium 3. Added to this package is the execution speed guaranteed by the frontier agent that introduced the first three models. Among them, Kiro Autonomous stands out for its support for the software development phase, leaving developers with almost only final control, flanked by security and DevOps agents to facilitate the “go-live” of billions of agents in the enterprise. Don't forget the innovation of Transform, the bridge that shortens your transition. It automates rewrites and migrations, reduces technical debt, and frees teams from the constraints of legacy systems. As Swami Sivasubramanian, Vice President of Agentic AI at Aws, points out, “A useful agent is not one that can do everything, but one that can be trusted because it operates within clear boundaries.”
