Microsoft agent framework explainer

Microsoft is drawing a sharper line between two parts of enterprise AI agents: the code developers write and the managed service that runs it in production. (youtube.com) An agent is software that uses a large language model to decide what to do next, call tools, and work through a task over several steps instead of just returning one chat reply. Microsoft says Foundry Agent Service is the hosted layer for building, deploying, and scaling those agents, with runtime features like identity, observability, and security controls. (learn.microsoft.com) The developer-facing layer is Microsoft Agent Framework, an open-source SDK and runtime that supports Python and.NET and can connect to models from Microsoft Foundry, OpenAI, Anthropic, Ollama, and others. Microsoft’s documentation says the framework includes agent sessions, middleware, memory hooks, tool integration through Model Context Protocol clients, and graph-based workflows for multi-step orchestration. (learn.microsoft.com) Microsoft’s own Foundry integration page spells out the split in two modes. A “Responses Agent” is code-first and does not create a server-managed resource, while a “Foundry Agent” is a versioned, server-managed agent defined in the Foundry portal or through service APIs. (learn.microsoft.com) That separation is showing up in Microsoft’s product packaging this week. In an April 22, 2026 blog post, the company said Microsoft Agent Framework reached v1.0, Foundry Toolkit for Visual Studio Code reached general availability, memory in Foundry Agent Service remained in preview, and a faster hosted-agents experience in Foundry Agent Service was also in preview. (devblogs.microsoft.com) The video published April 22 by Microsoft Mechanics frames the workflow as local development first, then deployment into Foundry for secure identity, scoped permissions, monitoring, publishing, and governance. Its chapter list runs from “Build agents at scale with Foundry” to “Manage performance,” with a sales-meeting-preparation demo in the middle. (youtube.com) Microsoft’s documentation now describes Foundry Agent Service as able to host prompt agents and code-based hosted agents built with Agent Framework, LangGraph, or custom code. The service page also lists managed tools such as web search, file search, memory, code interpreter, Model Context Protocol servers, and custom functions. (learn.microsoft.com) The company is also presenting the two options side by side in tutorials for existing apps. One Microsoft Learn guide says Agent Framework offers faster local performance and maximum code control, while Foundry Agent Service trades that for Azure-managed autoscaling, built-in content safety, built-in agent identity, authentication, and Microsoft 365 and Teams deployment paths. (learn.microsoft.com) Microsoft’s April 22 post ties the framework to earlier internal agent efforts, saying Agent Framework combines Semantic Kernel’s enterprise features with AutoGen’s multi-agent orchestration ideas. The result is a product stack that looks less like one monolithic “agent platform” and more like a familiar enterprise pattern: SDK on the developer side, managed runtime on the operations side. (devblogs.microsoft.com) That is the pitch in Microsoft’s latest explainer: build agents with regular software tools, then hand off execution, policy, and monitoring to Foundry when those agents need to run inside a company’s systems. (youtube.com)