Microsoft operates at enormous scale with Microsoft Foundry, where more than 80,000 enterprises build AI agents. Microsoft 365 Copilot alone serves over 20 million users with first-party agent usage growing 6x year-to-date. We spoke with Marco Casalaina, VP of Products for Microsoft Core AI, about lessons from running these systems in production.

The shift from chatbots to agents that do meaningful work — booking meetings, running analyses, sending emails — changes the engineering problem entirely. A wrong chatbot answer is a bad experience; a wrong agent action is a business incident. Prototypes fail in production because real users ask unanticipated questions, documents go stale, model updates change behavior subtly, and without identity controls or observability, there's no audit trail.

Marco's biggest lesson: "the harness matters as much as the model." The harness includes the inference layer (supporting 11,000+ swappable models), agent runtime for orchestration, and observability/governance via Foundry Control Plane. Models change constantly and require harness re-tuning — when Anthropic released Claude Opus 4.8, GitHub Copilot CLI had to re-tune before shipping. Framework neutrality lets agents run on any framework interchangeably.