Last week, AWS Vice President of Agentic AI Swami Sivasubramanian headlined re:Invent 2025 with a keynote focused on making AI agents easier to build, scale, and trust.

Key announcements included significant expansions to the Strands Agents SDK, now supporting TypeScript and edge devices to unlock new development possibilities. Amazon Bedrock AgentCore received a major upgrade with episodic memory functionality, enabling agents to understand user behavior and recognize patterns across interactions—making them smarter with every experience.

On the customization front, AWS introduced Reinforcement Fine-tuning in Amazon Bedrock for improved model accuracy, alongside new SageMaker AI tools that slash customization timelines from months to days. The company also announced checkpointless training on SageMaker HyperPod, allowing recovery from faults in minutes across thousands of AI accelerators.

Perhaps most notable was Amazon Nova Act, a new service for building and managing agent fleets that automate production UI workflows at scale. The platform offers end-to-end training for comprehensive agent interactions.

Sivasubramanian emphasized trust as paramount, with AWS implementing automated reasoning algorithms to logically verify agent behavior. Combined with updates to yesterday's Frontier Agents (Kuro, DevOps, and Security), AWS is positioning itself as the platform where agents work alongside humans most effectively.

In case you missed it, you can still revisit Dr. Sivasubramanian's full keynote at the AWS Events YouTube channel: