Microsoft Agent Framework merged Semantic Kernel and AutoGen into one SDK, and writing an agent became the easy part. Running it is where the effort goes: containers, session state, identity, scaling, monitoring. Since July, hosted agents in Microsoft Foundry Agent Service take over that layer for any framework, and classic agents retire in March 2027. What does the platform actually run for you, and what is still yours? How do you keep tool credentials out of agent code? What does a release pipeline for an agent look like, with versions, evaluations and rollback? And when is a hosted agent the wrong choice? We’ll take a real enterprise agent from local code to a production-ready hosted deployment and show where the platform helps and where it doesn’t.