On AI and Knowledge — Pablo Castro, Distinguished Engineer & CVP for AI Knowledge, Microsoft

AI-driven knowledge integration relies on three distinct pillars: intrinsic, extrinsic, and learned knowledge. Intrinsic knowledge, derived from model training, powers foundational capabilities like GitHub Copilot and accelerates software development. Extrinsic knowledge, facilitated by systems like Microsoft IQ, grounds agents in ambient organizational data—including documents, chat threads, and analytics—using advanced retrieval methods that combine vector and lexical search for superior accuracy. Finally, learned knowledge creates continuous improvement loops, where agents utilize the Agent Optimizer to evaluate performance, iterate on instructions, and refine configurations through automated hill-climbing. By layering these approaches within platforms like Microsoft Foundry, developers balance ease of use with granular control, enabling agents to participate effectively in complex organizational workflows while maintaining token efficiency and high-quality output.