Sivasubramanian highlighted real-world customer use cases—from ocean clean-up initiatives to brain-mapping applications—to demonstrate the versatility of AI agents that can proactively search, act and assist far beyond basic chatbots. But building such systems is challenging, he noted, requiring powerful models, robust code and integrated tools. “In this new world, we believe how you build agents should be really simple,” he said.
AWS introduced new capabilities for its Strands Agents SDK, including TypeScript support and deployment on Edge devices, aimed at accelerating the development of agentic systems. He acknowledged that moving agents from prototypes to production remains too complex for most companies.
To address that gap, AWS announced Amazon Bedrock AgentCore, designed to streamline deployment by managing the underlying orchestration while developers focus on innovation. Sivasubramanian also unveiled AgentCore Memory (episodic memory), enabling agents to understand user behaviour over time and recognize patterns across similar scenarios. “The more your agents experience, the smarter they become,” he said.
Additional announcements included Reinforcement Fine-Tuning in Amazon Bedrock for improved model accuracy and new model-customisation tools in Amazon SageMaker, cutting development timelines from months to days. AWS also introduced checkpointless training on SageMaker HyperPod, allowing large-scale training jobs to recover from faults within minutes—a change Sivasubramanian described as “a paradigm shift”.
AWS said its newly launched Amazon Nova Forge will further simplify enterprise AI adoption, offering a smoother path to build and scale advanced agentic systems.