AWS re:Invent 2025 in Las Vegas focuses on what the cloud provider itself sees as the next big step in AI transformation: ‘agentic AI’.
More than 60,000 attendees at the Venetian and millions of viewers online, even for the first time via Fortnite, are witnessing the ambitious vision AWS is presenting this year: the future of software is based on agents. CEO Matt Garman explains: “We are at a tipping point… AI assistants are giving way to AI agents that perform and automate tasks for you. This is where companies are finally starting to extract tangible value.”
This vision is by no means new: almost every technology company with a platform offering has been proclaiming the rise of agent-based AI for a year now. At re:Invent, AWS is trying to clarify what role it can and wants to play in that narrative.
AI Agents as a New Layer
The keynote by Dr. Swami Sivasubramanian, VP Agentic AI at AWS, forms the substantive core of the event. He illustrates the evolution from classic chatbots to fully autonomous agents with an example: a chatbot tells you what to investigate if your website visitors drop due to a bug, while agents actually perform tasks.
“AI agents can actively conduct research, consult data, detect errors, and propose and implement solutions.” According to him, this gives developers a sense of freedom: to build something that no longer just reacts, but acts autonomously.
However, few projects reach production. Garman acknowledges this, noting that companies often build beautiful prototypes that “never cross the finish line” because they are not secure, controllable, or scalable enough. Martin Elwin, Technology Director at AWS, also states that almost all CIOs are engaged with AI, but the majority are not yet realizing the value they expected. According to him, this is because they often start too technically or don’t think from the perspective of the problems they are experiencing.
Bedrock Agent Core: From Development to Production
AWS is trying to solve this with Bedrock Agent Core, a platform designed to make building, securing, and monitoring AI agents easy. Swami emphasizes that an agent only truly has value when it does more than just clever reasoning. An agent must have memory, secure identity access, and reliable observability capabilities. Only then can an agent make correct decisions autonomously.
He warns that “developers cannot solve what they cannot see”: without those monitoring, debugging, and testing capabilities, an agent remains unpredictable. Therefore, agents must be extensively evaluated, tested, and monitored before they go to customers.
AI doesn’t bring the expected value for some CIO’s, because they start out too technical.
Martin Elwin, Technology Director AWS
The new episodic memory plays a significant role here: agents must not only remember what happened before, but also why it is relevant. Swami gives a simple example from his own life: when he travels alone, he plans his departure to the airport differently than when he travels with his children. An agent must therefore not only remember that it is booking a flight, but also under what circumstances: solo or with family, in a hurry or not, light or heavy travel.
A good agent automatically recognizes that context and adjusts its choices accordingly. This makes the difference between a system that merely regurgitates information and an agent that truly understands what the user is trying to do.
read also
AWS announces new AI innovations: Nova, Nova Forge, and Novellas
Additionally, AWS announces new capabilities:
- Agent Core Policies to guide behavior without completely restricting autonomy.
- Agent Core Evaluations, allowing companies to simulate thousands of scenarios to detect drift and unexpected behavior before an agent goes live.
AWS Distinguished Engineer Marc Brooker emphasizes that successful agents not only need good prompts, but it’s primarily about carefully selecting tools, clear policies, and reliable interfaces. Too many tools or poorly described actions make agents unpredictable.
New Infrastructure
AI agents require computing power on an unprecedented scale. This year, AWS is heavily investing in infrastructure innovation. Trainium3 is AWS’s latest chip that powers the Ultraservers, AWS servers for training AI. Garman states that the largest configurations combine 144 chips into a single compute core with hundreds of petaflops of power. He asserts that AWS is thereby enabling a new category of agentic workloads “that you won’t find anywhere else”.
The chip is immediately available to customers and is already being intensively used by AWS itself for running large-scale Bedrock models.
AI Factory: AI Power in Your Own Data Center
AWS also introduces the AI Factory, a type of AWS-managed stack that companies can run in their own data center, including Trainium clusters, Nvidia GPUs, and services like Bedrock and SageMaker. According to Garman, the AI Factory primarily addresses the compliance and sovereignty needs of European companies: they want to use AI, but only if they have control over their data.
We want their data to reside where customers choose.
Martin Elwin, AWS Technology Director Northern Europe, confirms the demand for sovereignty based on conversations with customers: digital sovereignty and data location are top priorities, especially in the Benelux. He points to the combination of the Nitro architecture and the upcoming European Sovereign Cloud as a response to these demands.
A small caveat here is the fact that American cloud providers can never truly offer complete sovereignty. They are, after all, bound by the Cloud AI Act, which allows the US to compel AWS to transfer data, even when it resides in a European AWS data center.
New Frontier Models
A key component of AWS’s AI strategy is new models. Therefore, the Nova model family is being expanded. According to AWS, Nova offers strong price-performance ratios, low latency, and multimodal capabilities. Nova 2 will have different variants for reasoning, conversation, and efficient inference. The models are integrated into Bedrock and optimized for Amazon’s own hardware, such as Trainium.
To make the models even more accessible, AWS is also launching Nova Forge, a platform that provides access to checkpoints of the Nova models. These are saved versions of a model at a specific point in the training process. They often contain parameters and additional information. Additionally, there is Transform, a set of tools that allows companies to easily convert their own data into custom agents. This helps companies modernize existing workloads and make them ready for use more quickly.
Model Choice Becomes a Strategic Advantage
AWS emphasizes that there will never be one model that solves all problems. Garman says that customers, on the contrary, want to combine different models to make their agents more flexible. Bedrock supports a wide range of models, including Meta Llama, Mistral, Nvidia, and Amazon Nova.
Hugging Face Product Director Jeff Boudier sees the same trend. He calls AWS the “most open” of the major cloud providers when it comes to supporting thousands of open models and fostering choice for businesses. According to him, customers increasingly want to combine models for different tasks: “there isn’t one model that can do everything”.
AI Agents Are the Future
At re:Invent 2025, AWS positions itself as the architect of a new software era. The cloud is no longer the final destination, but is the foundational layer for autonomous AI systems that reason, remember, and execute. With Bedrock Agent Core, Trainium3, Nova, and the AI Factory, AWS provides a platform on which companies can not only build agents, but also (finally) deploy them securely and scalably.