AWS (Amazon Web Services) is multiplying new features at
re:Invent , the 2025 edition marks a clear shift:
the future lies in generative AI , its concrete uses and above all its agents capable of acting alone.
A turning point that Stephan Hadinger , Director of Technology at AWS France, detailed with a very operational vision during an interview , with the aim of accelerating application modernization, strengthening model governance and preparing the company for a world where autonomous agents continuously monitor systems…
Modernize to better leverage AI
Despite the maturity of the cloud, the majority of global applications remain ” on-premise “, that is to say, on site.
This observation has led AWS to introduce AWS Transform , a suite of tools designed to audit and modernize environments still burdened by technical debt. The goal of these tools is to free up time and capacity for data, generative AI, and agentic AI projects.
AWS’s philosophy remains unchanged, and the company still aims to offer choice, particularly through Amazon Bedrock , which now brings together nearly 250 models from 34 suppliers. This dense ecosystem requires support and a structured approach.
For Stephan Hadinger, the criteria remain unchanged: performance, cost, and latency. The challenge now is to help companies select not the model that ” does everything ,” but the one that corresponds to a specific use case .
Governance, transparency and security: the triple requirement
With the increasing use of agents, AWS is emphasizing oversight . The new Agent Policy component acts as an independent safeguard, validating each action before execution. This mechanism is designed to prevent unforeseen behavior and provide complete traceability.
In addition, there is continuous monitoring to detect performance deviations , as well as adjustable ” guardrails ” for each customer. For regulated sectors, such as finance, insurance, or HR, AWS offers an automated verification system for internal policies, which are transformed into formal rules applied to each interaction.
Finally, regarding privacy, Stephan Hadinger points out that neither AWS nor template providers have access to prompts or customer data. This approach is reinforced by the requirements of the AI Act , recently adopted by the European Parliament …
Are “border agents” the new key players in the system?
The other major new feature concerns Frontier Agents , with agents that wake up on their own, analyze logs, run tests, detect anomalies and prevent incidents before they even occur.
This generation also marks a break, because the agents are no longer simple executors , but active monitoring systems, close to permanent control of the development and production cycle.
AWS is strengthening its entire ecosystem…
Alongside its software announcements, AWS unveiled Trainium3 , a new training chip boasting up to 4x performance gains and a 40% reduction in power consumption. The UltraServer, built around this chip, paves the way for a future Trainium4 compatible with Nvidia GPUs.
The Nova family of models is also expanding, accompanied by Nova Forge , a platform for refining models on proprietary data. For organizations seeking to maintain control of their infrastructure, AWS is finally launching its AI Factories , veritable AI factories deployable in private data centers, in partnership with Nvidia.