What happens in Las Vegas ... Nah - let's have a look. All things (Gen) AI. Re:Invent 2025 edition
written by Stefan Christoph
What happens in Las Vegas … Nah - let’s have a look. All things (Gen) AI. Re:Invent 2025 edition
What happens in Las Vegas … Nah - let’s have a look. All things (Gen) AI. Re:Invent 2025 edition
Created on 2025-12-03 10:34
Published on 2025-12-03 10:49
From the original 2024 article:
I turned out to be a very challenging but good learning experience. Although is it pretty hard to follow all those updates - even just on one single topic like GenAI. For me it turned out to be useful way of absorbing the information, I hope you as a ready get some benefit from this as well.
Disclaimer: I’m writing this over my morning coffee and didn’t had the time to dive in every detail of the new announced features and service. I’m just taking you with me on me personal discovery of the new announcements and hope this turns out useful."
In this document you find
a daily summary for each of the event days and - more importantly -
a breakdown of the releases in functional areas.
Daily Summary
Wednesday
Amazing keynote by Swami Sivasubramanian. At least the first 40 minutes are super useful, even if you are not super interested in new service launches. Swami nicely breaks down how AI agents and Agentic AI are relevant, how they work and what are common challenges to adopt them. Being at an AWS keynote, he obviously also explain how you can implement your own Agents successfully.
William Brennan,Blue Origin talked intensively about their journey to take systems build by AI Agents, into space. Well, not everyone flies into space, but a lot of Blue Origin’s insights to be adopted to any enterprises. Some key aspects I noted are 1/ that they where able to extend the team of experts with a much larger number of AI Agents driving productivity, 2/ AI adoption is not just a technology job, 3/ you need to provide the right building blocks for teams to master the innovation process, 4/ you need to build on top of well established primitives to get rid of the undifferentiated heavy lifting, and 5/ super important: Blue origin is now working on making more of their data accessible and actionable for AI agents to be able to accelerate their Agentic AI project even more.
So overall - I still have 1h to go - super impressive keynote. Less focussing on a huge number of announcements( “just” 10 for the day) but really focussing on putting things into perspective - loved it!
Tuesday
Screenshot from the keynote
Second day’s keynote by Matt Garman was packed with many, many (44 overall) significant announcements. Trying below to digest and categorize what I have seen in the post below. Let me know if this is helpful. And yes, Mat also addressed vocally self-critical that also for AWS naming is a very hard problem. I suggest to have a watch of the keynote in [8] to answer the “Trainium for Inference?” question for yourself ;) 
Tuesday’s announcements from [10]
Release breakdown by functional area
New releases which help you to excel the building of your Agents:
Amazon Bedrock AgentCore Runtime now supports bi-directional streaming. This is awesome change. I guess we all have been experience the frustration, of prompting an agent to do something, but then while the agent picks up the work and plays back some intermediate results we figure - wait. That goes in the wrong direction. I want to steer into a slightly different direction, without breaking the entire interaction and restart it. This is now possible to build with Amazon Bedrock AgentCore. Dive deep into this in [1]
**Amazon Bedrock AgentCore **now includes Policy (preview). This addresses a concern many customer raised with me and to be quite frankly - a concern of mine as well. While I’m happily delegate tasks to an agent and love the agent to display agency to the extend to fulfill the task by it’s own without asking to many follow ups and permission to do certain things, I would need to build trust that the agent is not going to do things it shouldn’t do. With the policy feature in AgentCore builders can not build agents with defined policies attached setting clear boundaries for the agents. Dive deep into this in [1].
Amazon Bedrock AgentCore now includes Evaluations (preview). How well are you agents doing? Prior for putting them to production we need to have a good idea about this. And then also while running in Production - how are they doing facing real customer input? With the new Evaluation feature builder can monitor their agent which ultimately leads to more and better agents for customers. Dive deep into this in [1].
Amazon Nova Act, a new AWS service for building and managing agents that automate production UI workflows, is now General Available. Super helpful to e.g. automate repetitive UI Workflows. Dive deep into this in [7]. US only though.
Strands Agents SDK, an open source python framework that takes a model-driven approach to building and running AI agents in just a few lines of code. While this is super useful to customers and AWS as well, python is not everyone’s pice of cake,. Now you can also use TypeScript to build your agent with the Strands Agents SDK. Dive deeper in [12].
New Frontier Agents which are here to support you, even when you’re not around:
AWS Security Agent proactively securing your application in the SDLC - you just go another security person in your team! Dive deep into this in [2].
AWS DevOps Agent (preview) Â autonomously resolves and prevents operational incidents across AWS, multicloud, and hybrid environments. Dive deep into this in [6].
Kiro autonomous agent: The frontier agent for software development
New Models to fuel your Generative AI applications and AI Agents.
Not one model for every use cases but a model for each use case is important to build performant and cost-effective solution for your customer’s demand. Amazon Bedrock just put a “few” - actually quite some - at your finger tips.
18 additional fully Managed open-weight models , which are including the new Mistral Large 3 and Minstral 3 family of models have been just added to this extensive menu of models. More pain of choice for builders, but many more good applications and agents going to be build for happy customers :) Dive deep into this in [3].
**Amazon Nova 2 family **with
Nova 2 Lite, a cost-effective reasoning model,
Nova 2 Pro(preview) for the more challenging tasks,
Nova 2 Omni(preview), all-in-one model supporting multimodal reasoning and image generation and
**Nova 2 Sonic **the new speech -to-speech model
has been announced. Still need to build experiences with this, but this looks very exciting! Dive deep into this in [4].
Build your own Frontier Models
Build your own frontier model has often been a very, very bad advise for customers. Huge investments required and terrible technical challenges to be overcome.
This changed. It might be now a much, much better advise for customers. Don’t get me wrong. By far not for your everyday use case, but now it became very accessible for use cases which matter.
With Amazon Nova Forge, you can build you own models based on Amazon Nova model. On a very high level you can pick up Nova Models in various stages of their training and continue from there with your own training, based on your own data. With a great support of Amazon Nova Forge, which eases technical and data challenges for your own training. Super exciting and you can dive deep into this in [5]
Fine-tune off-the-shelf models to laser-focus on your own use case
Reinforcement fine-tuning support in Amazon Bedrock enables developers to improve model accuracy without extensive ML expertise or large labeled datasets. As Swami explains in his keynote, models are the driver of AI agents. Without the ability to fine-tune models, we often end up using very large models with very strong and generic capabilities, which also - hopefully - cover the capabilities we actually need. This is often times expensive and waste of resources. With this support, fine-tuning becomes way more accessible. Today for Nova Models, but will be extended to open-weight models. Dive deeper in [11].
How to participate and learn from home?!
But re:invent is so much more than “just” announcements and new services&features. The real value for myself and many builders out there are the fantastic session which allow us to learn about new things and learn from others. Good thing, quite some of this year’s breakout session are already available at YouTube [9]. Have fun!
References
[3] https://aws.amazon.com/blogs/aws/amazon-bedrock-adds-fully-managed-open-weight-models/
[4] https://aws.amazon.com/nova/models/
[7] https://nova.amazon.com/act
[8] https://www.youtube.com/watch?v=q3Sb9PemsSo
[9] https://www.youtube.com/playlist?list=PL2yQDdvlhXf9gdFFBcDPUHAJS7kkIkIet
[10] https://aws-news.com/?hideRegionalExpansions=true&types=news&startDate=2025-12-02&endDate=2025-12-02
[12] https://github.com/strands-agents/sdk-typescript
📝 Last updated: August 14, 2026 — Restored the article body, which was lost in the LinkedIn migration