Microsoft's Build 2026 announcement focuses on their Agent Harness and Foundry Hosted Agents, now in General Availability. This signifies a shift in how AI agents are perceived and built for production. Research indicates that approximately 98.4% of an agent system is infrastructure, not the AI model itself. The Agent Harness provides this crucial operational core, handling tasks like function invocation, context management, and tool routing. Foundry Hosted Agents offers this harness as a managed, consumption-billed service, simplifying deployment.The Agent Harness addresses common production bottlenecks like wiring agents to tools and persisting conversation history. Microsoft's core bet is that this harness infrastructure, not the interchangeable AI model, is the true product. The harness includes features like function invocation with history, context compaction, a plan-and-execute todo list, file memory, and built-in OpenTelemetry for observability. This ensures that every action, from tool calls to approval decisions, is traceable.Foundry Hosted Agents eliminates the need for complex YAML configurations by providing a managed deployment where users supply only the chat client, instructions, and tools. This managed service shares the same core logic as the locally runnable harness, ensuring consistent behavior across development and production environments. The framework also introduces connectors for GitHub Copilot and the Claude Agent SDK, allowing for model swaps without altering the harness. This unified governance policy plane ensures consistent approval rules and audit trails across different underlying AI models. The durable layer of the agent system, encompassing approval gates, history, and policy enforcement, is highlighted as the key area for engineering investment. Microsoft's offering positions the harness as a stable, supported product, enabling developers to focus on building adaptable applications.