AWS pushes AI agents: from generative models to guided action in business processes

What is the difference between a generative AI model and an AI agent? The former generates text, code or images based on a prompt. The second takes a goal, plans, uses tools, reads and writes to business applications, executes and verifies, all to achieve that goal. It is this gap - from content generation to focused action - that Amazon Web Services has put at the heart of its strategy: no longer showcases of templates, but platforms and tools to build customised agents that enterprises can train on their data, govern deterministically and put into production quickly.

"I believe the advent of AI agents has brought us to a turning point in the trajectory of artificial intelligence: from technical marvel to real value. Agents are exciting because they can take action and get things done,' said CEO Matt Garman in his keynote that opened re:Invent 2025, the annual event that outlines strategy and news from Amazon's cloud division. Which this year is playing all its cards on artificial intelligence or, better, on agentic AI, with the creation of a complete and functional ecosystem to allow companies to set up their agentic systems simply and quickly, without having to deal with infrastructures and models: a ready-to-use SDK that for Aws means involving and keeping companies, large and small, hooked to its ecosystem.

The basis is the Bedrock AgentCore platform, which provides the 'building blocks' for building capable but governable agents: all the most advanced GenAI models are available, from ChatGPT to Anthrtopic's Claude, in which the group has invested, to the new models of the European Mistral, so that the user can focus on the application and use cases. Two automatic features are added at this stage: a policy feature that imposes limits on agent behaviour and an evaluation feature for monitoring them in the real world in terms of correctness, usefulness and security.

At the infrastructure level, the offering is enriched with Ai Factories, aimed at bringing Aws services into customers' existing data centres with their proprietary data assets, respecting sovereignty, policy and governance. At the same time, Garman announced the launch of the new Trainium 3 chip with twice the energy efficiency of version 3, aimed at large-scale training and global inference, a product that formally aims to integrate with Nvidia's family but at a strategic level seems to want to reduce dependence on the AI chip giant. Aws's agentica platform is based on the models of the new Nova 2 family, also enriched by Omni, a unified multimodal model capable of integrating speech, text, video and images. Building on these models is also Nova Forge, the service enabling the creation of customised frontier models.

The end point of the vision is indeed the frontier agent class: autonomous, scalable and long-running agents that can work alongside teams, orchestrating heterogeneous tools and data without continuous intervention. This is where the 'agent Sdk' becomes a competitive advantage: policy, evaluation and memory features give confidence, AgentCore guarantees elasticity, Trainium 3 power and training capabilities. Added to this package is the speed of execution guaranteed by the frontier agents, of which the first three models were presented, among which Kiro Autonomous stands out for supporting the software development phases, leaving developers almost only the final control, flanked by the Security agent and the DevOps agent, which facilitate the 'putting to work' of billions of agents in companies. One cannot forget the innovation of Transform, the bridge that shortens the transition: it automates rewrites and migrations, reduces technical debt and frees teams from the constraints of legacy systems. As Swami Sivasubramanian, vice president Agentic AI at Aws, points out, 'useful agents are not those that can do everything, but those that are reliable because they operate within clear boundaries'.