AWS Summit New York City 2025 Keynote | Amazon Web Services VP of Agentic AI Swami Sivasubramanian

Sivasubramanian explains the fundamental differences between agentic applications and traditional software. He outlines three key differences: decomposition (agents take high-level objectives and break them down into plans and code), self-reflection (agents evaluate their progress and adjust tactics as needed rather than executing linearly), and action/tool use (agents can perform actions in other systems and access specialized tools). This section highlights the need for new tools and frameworks for building agentic software. AWS has learned that customers want more control and the ability to use open frameworks with AWS services. Sivasubramanian introduces Strands Agents SDK, an open-source framework that simplifies agent development by embracing the capabilities of state-of-the-art reasoning models. He also discusses the importance of emerging standards like MCP (Message Connection Protocol) and A2A (Agent-to-Agent) for interoperability, with AWS serving on the steering committees for both projects. Sivasubramanian identifies the key challenges in deploying agents to production, focusing on their dynamic and autonomous nature. He outlines several critical requirements: secure execution and scaling of agent code, memory management for past interactions and learning, identity and access control for autonomous interaction, tools for tasks like code execution and browser interaction, discovery and connection mechanisms for tools, and deep auditability of agent actions. He emphasizes that these challenges create significant friction when moving from proof-of-concept to production deployment, especially when business needs change or new models and frameworks emerge. The section establishes the problem that the subsequent announcement of Amazon Bedrock AgentCore aims to solve, highlighting how these technical challenges have slowed down customer adoption of agents in production environments.