AWS re:Invent 2025 - Keynote with Dr. Swami Sivasubramanian
In this keynote, Dr. Swami Sivasubramanian, Vice President of Agentic AI at AWS, discusses the transformation of AI agents and announces several key innovations that lower technical barriers and accelerate development from idea to impact (4:37-4:45).
The keynote emphasizes the general benefits of Agentic AI (1:03-1:06), highlighting its ability to make decisions and deliver beyond imagination, empowering scientists and engineers. It contrasts agents with chatbots, explaining that agents investigate, diagnose, and initiate solutions, rather than just providing advice (7:18-7:29). The core components of an agent are identified as a model, code, and tools (7:57-8:39).
Key announcements and takeaways from the video include:
Strands Agent SDK - Support for TypeScript and Edge Devices/Robotics (11:39)
Announcement: Strands Agent SDK, an open-source tool, now supports TypeScript and edge devices, including robotics.
Takeaway: This extends Strands to one of the world's most popular programming languages and enables autonomous AI agents to run at the edge, unlocking new capabilities in automotive, gaming, and robotics. It aims to simplify agent building with minimal code and improve accuracy and code maintenance.
Amazon Bedrock AgentCore - Episodic Memory Functionality (21:12)
Announcement: AgentCore's long-term memory now includes Episodic Memory, allowing agents to remember and learn from past experiences.
Takeaway: This feature enables agents to truly understand user behavior, adapt automatically by recognizing patterns across similar situations, and proactively offer suitable solutions. The more an agent experiences, the smarter it becomes by recalling specific interactions as discrete episodes.
Amazon Bedrock - Reinforcement Fine Tuning (RFT) (39:05)
Announcement: Bedrock now automates the entire Reinforcement Fine Tuning (RFT) workflow, making this advanced model customization technique accessible.
Takeaway: RFT allows models to learn from the outcomes of their actions, achieving significant accuracy gains (e.g., 66% on average over base models). Bedrock simplifies the process by handling the complexity of reinforcement learning implementation.
SageMaker AI - Serverless Model Customization Capabilities (43:48)
Announcement: New serverless model customization capabilities have been released in SageMaker AI.
Takeaway: This allows customization of popular models like Amazon Nova, Quinn, Llama, and Deepseek for deployment on Bedrock or SageMaker in just a few steps. It offers both a self-guided approach and an agent-driven experience that uses an AI expert to guide the full customization workflow, reducing the process from months to days.
SageMaker HyperPod - Checkpointless Training (50:03)
Announcement: Checkpointless training is now available on SageMaker HyperPod.
Takeaway: This new feature significantly reduces recovery overhead by continuously preserving the model state across the distributed cluster. It allows for automatic recovery from infrastructure failures in minutes with zero manual intervention, even with large clusters, leading to faster recovery and significant cost savings in model training.