re:Invent Highlights 2025

Oh What a Year for Agentic AI! AWS re:Invent showcased major advances in how AWS is shaping the next chapter of AI infrastructure, with a strong focus on frontier agents and autonomous systems.

Let's Talk Highlights

The keynote highlighted three new frontier agents that operate like true teammates, acting predictively, following intent, and respecting clearly defined boundaries. This is powered by AgentCore, AWS’s modular framework for managing agent behavior. AgentCore adds stronger policy controls and introduces real world behavioral evaluations, solving a long standing challenge in assessing agent performance during live operation. AWS also introduced native agents including Kiro, the AWS Security Agent for continuous code scanning and on demand pen testing, and the AWS DevOps Agent, which identifies incidents, traces root causes, and explains how to fix and prevent future issues.

A central theme was AI infrastructure at massive scale, including the announcement of new P6e GPU instances. The keynote also highlighted Humain and the AI Zone in the Kingdom of Saudi Arabia, areas where Atos plans to expand. AWS expanded its AI Factory vision with customer specific, rapidly deployed AI environments fully managed by AWS. These factories integrate with existing enterprise systems and leverage the new Trainium 3, alongside the milestone of more than one million Trainium chips already in use.

On the inference side, AWS introduced new Bedrock models from Google, Nvidia, and Mistral AI, including two Mistral Mini models. The Nova family expanded with Nova 2 Lite, Nova 2 Pro, and Nova 2 Sonic. Combined with Bedrock Knowledge Bases, these updates enable rich multimodal workflows and allow organizations to pair their data with deep domain expertise for more precise AI results.

A big push forward was Nova Forge, a platform for creating custom frontier class models, even capable of generating a full novella. Nova Forge supports open training and post training workflows, and models built within it can be imported directly into Bedrock for inference. This approach, which merges enterprise knowledge with advanced training pipelines, will resonate with companies building specialized AI systems.

We also heard about updates to Amazon Quick Suite and Amazon Connect, which now powers a leading agentic cloud contact center exceeding one billion dollars in business. AWS also announced major improvements across storage and databases, including 50TB S3 object support, intelligent tiering for S3 tables, and new Database Savings Plans. Together these innovations reflect AWS’s commitment to accelerating developer velocity, strengthening security, and enabling organizations to operate scalable autonomous AI systems.

What were my favorites?

AgentCore Memory (with Episodic Functionality):

This service enhances AI agent capabilities by giving them the power to learn from past experiences, similar to human episodic memory. It allows agents to remember specific past interactions and their outcomes, which improves decision-making over time and provides more tailored insights to users. This feature lifts the burden of building custom memory systems off developers. AgentCore is fully managed, scalable, and includes enterprise-grade security and encryption, making it ideal for large-scale, production-ready AI agent systems. It supports advanced agent use cases by enabling learning and adaptation across sessions.

Model Customization in Amazon SageMaker AI:

This serverless capability empowers developers to customize popular foundation models quickly, reducing the end-to-end workflow time from months to days. It offers a unified, easy-to-use interface and supports the broadest set of advanced techniques, including Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Reinforcement Learning from AI Feedback (RLAIF). The service is entirely serverless, automatically handling compute provisioning, scaling, and optimization. It includes an AI agent-guided workflow (in preview) to help define use cases, generate synthetic data, and evaluate models for specialization like brand voice alignment or domain expertise.

Checkpointless Training on Amazon SageMaker HyperPod:

Checkpointless training is a new foundational model training feature designed for fault-tolerant, large-scale Generative AI development. It fundamentally shifts from traditional checkpoint-based recovery, maintaining continuous forward training momentum even when infrastructure faults occur. By automatically swapping faulty components and using peer-to-peer state transfer from healthy accelerators, it reduces recovery time from hours to just minutes. This capability helps organizations achieve over 95% training goodput on clusters with thousands of AI accelerators, saving millions in compute costs by eliminating idle accelerator time.

Reinforcement Fine-tuning (RFT) in Amazon Bedrock:

RFT simplifies and automates the advanced model customization technique of reinforcement learning, making it accessible to everyday developers without deep machine learning expertise. The process trains models using a small set of prompts and a feedback-driven approach, which eliminates the need for large, labeled datasets. RFT improves model accuracy by 66% on average over base models, enabling the use of smaller, faster, and more cost-effective model variants while maintaining high quality. Developers use reward functions (rule-based or AI-based judges) to define what constitutes a good response for their specific business needs.

Looking forward to 2026

AI Agents and Infrastructure: AWS re:Invent showcased major advances in how AWS is shaping the next chapter of AI infrastructure, with a strong focus on frontier agents and autonomous systems. The keynote highlighted three new frontier agents that operate like true teammates, acting predictively, following intent, and respecting clearly defined boundaries. This capability is powered by AgentCore, AWS’s modular framework for managing agent behavior. AgentCore adds stronger policy controls and introduces real-world behavioral evaluations, solving a long-standing challenge in assessing agent performance during live operation. AWS also introduced native agents, including Kiro, the AWS Security Agent for continuous code scanning and on-demand pen testing, and the AWS DevOps Agent, which identifies incidents, traces root causes, and explains how to fix and prevent future issues.

BIG PUSH COMING FOR GPUs! A central theme was AI infrastructure at massive scale, including the announcement of new P6e GPU instances. The keynote also highlighted Humain and the AI Zone in the Kingdom of Saudi Arabia, areas where Atos plans to expand. AWS expanded its AI Factory vision with customer-specific, rapidly deployed AI environments fully managed by AWS. These factories integrate with existing enterprise systems and leverage the new Trainium 3, alongside the milestone of more than one million Trainium chips already in use. On the inference side, AWS introduced new Bedrock models from Google, Nvidia, and Mistral AI, including two Mistral Mini models. The Nova family expanded with Nova 2 Lite, Nova 2 Pro, and Nova 2 Sonic. Combined with Bedrock Knowledge Bases, these updates enable rich multimodal workflows and allow organizations to pair their data with deep domain expertise for more precise AI results.

If you work for a partner, you will be seeing a surge with Nova Forge, which supports open training and post-training workflows, and models built within it can be imported directly into Bedrock for inference. This approach, which merges enterprise knowledge with advanced training pipelines, will resonate with companies building specialized AI systems. Additional highlights included updates to Amazon Quick and Amazon Connect, which now powers a leading agentic cloud contact center exceeding one billion dollars in business.

Time to Optimize how we deliver FinOps and Storage: Major improvements were announced across storage and databases, alongside new financial offerings. The maximum S3 object size was increased tenfold from 5TB to 50TB, simplifying workflows for massive files like high-resolution video and large AI datasets. S3 object support also received updates, including Intelligent-Tiering for S3 Tables, which automatically moves table data between access tiers to reduce storage costs without manual configuration. On the database front, AWS introduced Database Savings Plans, a major FinOps update offering up to 35% savings on eligible database usage (including Amazon Aurora, Amazon RDS, and Amazon DynamoDB) in exchange for a one-year commitment to a consistent hourly spend. This plan offers flexibility across database engines and regions. Furthermore, Amazon S3 Vectors was introduced and made Generally Available (GA), functioning as a cost-optimized vector storage and indexing service within S3 designed to drastically reduce costs for vector-based workloads.

The future of Data and Analytics with Quick Suite and SageMaker: The data and analytics portfolio saw significant enhancements, particularly integrating with the new AI capabilities. The next generation of Amazon SageMaker was positioned as the "center for all your data, analytics, and AI," with a Unified Studio experience integrating data engineering, analytics, and generative AI workflows. Amazon Q, the generative AI-powered assistant, received updates in QuickSight, including scenario analysis, which helps business users solve complex questions faster by guiding them through in-depth data analysis. Amazon Q in QuickSight now also generally supports unstructured insights, allowing it to augment analysis from traditional BI data sources with contextual information from unstructured sources like documents and emails. Together, these innovations reflect AWS’s commitment to accelerating developer velocity, strengthening security, and enabling organizations to operate scalable autonomous AI systems.

Any other questions, please contact me at info@aiaz.net or check out the official AWS re:Invent announcements here: https://aws.amazon.com/blogs/aws/top-announcements-of-aws-reinvent-2025/