My honest take on AWS re:Invent announcements

Dr. Ryan Ries here. I just spent the last week at AWS re:Invent and it was an amazing time. Every year AWS takes it up a notch and this year was no exception. 

Today, I’ll be covering my take on the announcements that came out this week. 

But first, I want to give a couple of quick shoutouts. 

First to the Mission team for all their hard work this week at re:Invent. All your dedication definitely paid off. 

Also, to our customers and friends at AWS. I loved having the opportunity to meet many of you this week. If I didn’t see you and you’d like to chat about your AI initiatives, let me know. I’m all ears. (Well, once my hearing comes back to normal after Chainsmokers). 

Agent Overload

Every presentation. Every booth. Every conversation. Agents, agents, agents.

At one point, I literally texted my team: “agents agents agents, so much spin.”

AI agents are a genuine shift in how we think about automation and delegation. I am just as excited about the possibilities for customers with Agents. But somewhere between the vision and the vendor pitch, we’ve lost the plot. 

Not everything is going to be an agent. There needs to be  a lot of thought and design decisions that need to be made to decide if you are going to build an agent or if you are better off just creating a really good prompt. 

Yes, it is easy to just say to have an agent do this or that, but in reality, you need to carefully consider your architecture and SLA’s to fit the system's needs.

I worked in the Augmented Reality space for a long time, and we got to a point where we used to talk about the “floating shoe” because that seemed to be all anyone would build. 

Here is an object in space you could look at, and people would be like, “Wow, that is amazing.” Now there’s this meme making the rounds that perfectly captures how I feel. 

(PS, if you’re not familiar with the floating shoe concept, it refers to basic applications that simply place static 3D models into a physical space, without animation, interaction, or a compelling use case.)

Agents are certainly worth your attention, and understanding the tooling for agents and what you can do with them is valuable, so let’s talk about that.

What Actually Caught My Attention

Buried underneath the agent hysteria, AWS announced some legitimately interesting developments.

Episodic Memory in AgentCore

An important part of many agentic systems is how you deal with short-term and long-term memory. AWS is building memory systems that attempt to understand your patterns over time. Not just “remember this fact,” but “learn how this person thinks and works.”

The technology mirrors what Simon Willison discovered when ChatGPT started adding Half Moon Bay signs to his pelican images without being asked. It had quietly built a profile of his life from scattered conversations. I touched on this in a previous Matrix.

Memory architecture in AI systems is a fundamental design question about how learning compounds over time. Get it wrong, and you bake irrelevant context into every future decision.

Get it right, and you might actually build something that improves with use rather than degrading into expensive noise.

Nova Forge and Real Customization

AWS introduced Nova Forge, which lets you access pre-trained, mid-trained, or post-trained models and customize them with your proprietary data. Four new Nova models were launched, three focused on text generation and one that handles both text and images.

The pitch is flexibility. You’re not stuck with whatever OpenAI or Anthropic trained on public internet data. You can take a foundation model and actually make it yours.

This matters for enterprises sitting on massive proprietary datasets that’ve been frustrated by the “one size fits all” approach of commercial APIs.

Serverless Model Customization in SageMaker

Speaking of customization, AWS launched serverless model building in SageMaker. You can now start training a model without thinking about compute resources or infrastructure. The system handles provisioning automatically.

They also rolled out Reinforcement Fine-Tuning in Bedrock. Choose a preset workflow or reward system, and Bedrock runs the customization process from start to finish. It’s SageMaker Autopilot for fine-tuning.

I’m cautiously optimistic about this and how it will help experienced data scientists and ML engineers quickly run tests to validate whether a fine-tuned model will improve results or if they are better served just staying with a foundation model.. We’ve spent years telling customers “you need a team or a partner with a team (like Mission) of ML engineers to productionize models.” Maybe we were wrong. Maybe the tooling can finally catch up to the ambition.

AI Factories and Data Sovereignty

AWS announced “AI Factories” in partnership with Nvidia. These are turnkey systems that let corporations and governments run AWS AI infrastructure in their own data centers.

You can stock it with Nvidia GPUs or opt for Amazon’s Trainium3 chip. The pitch addresses data sovereignty, the need for organizations to control their data completely without sharing it, even to use AI.

Trainium3 promises 4x performance gains for training and inference while cutting energy consumption by 40%. Andy Jassy said on X that Trainium2 is already generating revenue, which suggests Amazon is serious about competing with Nvidia’s stranglehold on AI compute.

The Stuff That Actually Saves Money

Tucked into dozens of announcements was something that made me smile: Database Savings Plans.

Commit to consistent database usage over a year, and AWS will cut your costs by up to 35%. Corey Quinn, chief cloud economist at Duckbill, summed it up perfectly: “Six years of complaining finally pays off.”

This is the kind of announcement that will make finance teams actually like you.

My Take on All of This

Two parallel tracks have seemed to emerge from re:Invent.

Track one is the agent hype cycle. Every vendor is racing to proclaim they’ve solved autonomous AI, often by rebranding existing automation tools and slapping “agent” on the label.

Track two is the quiet, unglamorous work of making AI infrastructure actually usable at scale. Better memory systems. Simpler customization workflows. Cost controls that make sense. 

Guess which one matters more for your business in 2025?

If you’re building AI systems right now, here’s what I’d focus on:

Memory architecture: How does your system decide what to remember and what to forget? What prevents casual conversations from corrupting business-critical decisions?

AgentCore Policy: I didn’t specifically talk about this one, but it is something to keep in mind, and something I was really glad to hear. Policy in Amazon Bedrock AgentCore actively blocks unauthorized agent actions through real-time, deterministic controls that operate outside of the agent code. You can read more about AgentCore Policy in this blog post from AWS.

Model customization: Can you actually train on your proprietary data, or are you stuck with generic models that know everything about the internet and nothing about your business?

Cost structures: Are you paying for peak capacity all the time, or have you locked in savings plans that match your actual usage patterns?

I’m curious what you’re seeing in your own organizations.

Let me know. I learn as much from your experiences as you hopefully do from mine.

Until next time,
Ryan

Now, time for this week’s AI-generated image and the prompt I used to create it:

Generate an image of a muppet at a Chainsmokers concert. The muppet is wearing a Mission T-shirt and he’s having a great time. You can see the pyrotechnics near the stage and the concert is at Drai’s in Las Vegas.