Microsoft’s Agent Orchestration
- A recent YouTube explainer compared Microsoft's Foundry Agent Service with the Microsoft Agent Framework for enterprises. - The discussion emphasised separating model choice, agent logic, and runtime deployment for governance and scaling. - Creators treating orchestration as infrastructure suggests buyers now ask how agents integrate with identity, security and observability. (youtube.com)
Microsoft is drawing a sharper line between writing an AI agent and running one in production, with Agent Framework 1.0 for code and Foundry Agent Service for managed deployment. (devblogs.microsoft.com) Microsoft said on April 3, 2026 that Agent Framework reached version 1.0 for.NET and Python, with support for multi-agent orchestration, multiple model providers, and interoperability through Agent-to-Agent and Model Context Protocol standards. (devblogs.microsoft.com) Foundry Agent Service, documented by Microsoft Learn last week, is the hosted layer: a fully managed platform to build, deploy, and scale agents while using models from the Foundry catalog and outside frameworks. (learn.microsoft.com) In plain terms, Agent Framework is the toolkit for how an agent thinks and coordinates work, while Foundry Agent Service is the managed environment for where that agent runs, stores state, and is monitored. Microsoft’s own documentation splits those layers by supporting both direct model inference and service-managed agents. (learn.microsoft.com, learn.microsoft.com) That separation tracks a problem large companies hit quickly: the model, the workflow logic, and the runtime controls do not change on the same schedule. Microsoft’s framework supports providers including Microsoft Foundry, Anthropic, Azure OpenAI, OpenAI, and Ollama, while the service side adds managed hosting and scaling. (learn.microsoft.com, learn.microsoft.com) Microsoft has been building toward that split since at least October 1, 2025, when it introduced Agent Framework in preview as a way to unify pieces from Semantic Kernel and AutoGen into one open-source SDK. The company said then that developers were struggling with orchestration logic, model connections, and hosting infrastructure as separate jobs. (devblogs.microsoft.com, devblogs.microsoft.com) The current Foundry runtime documents the basic unit of operation in equally separate parts: an agent defines the model, instructions, and tools; a conversation keeps the history; a response is the generated output. That design lets teams swap pieces without rewriting the whole system. (learn.microsoft.com) Microsoft’s hosted-agents preview adds the operational controls enterprises usually ask for after a prototype works once: managed hosting for containerized agents, scaling, and observability. Microsoft says those controls are aimed at secure operation across more complex cloud environments. (learn.microsoft.com) The YouTube explainer tied those pieces together for buyers evaluating Microsoft’s stack, showing local deployment into Foundry Agent Service with secure identity, scoped permissions, and monitoring of agent interactions. The pitch was less about a single model and more about the plumbing around it. (youtube.com) That leaves Microsoft selling agent orchestration less as a chatbot feature and more as enterprise infrastructure: open-source code for the logic, managed services for the runtime, and standards-based links between them. The closer customers get to production, the more that division appears to be the product. (github.com, learn.microsoft.com, devblogs.microsoft.com)
