From Container to Conversation: Hosted Agents in Microsoft Foundry

About this event

You've built an agent. It works on your laptop. Now someone wants it in production — with per-user isolation, secrets that aren't baked into the image, traces you can actually read, and a bill that doesn't scale with your prototype's idle time. That's the gap Microsoft Foundry Hosted Agents closes. You keep your code and your framework — Microsoft Agent Framework, LangGraph, Semantic Kernel, or something you wrote yourself. Foundry takes the endpoint, the identity, the scaling, and the session state. You don't get to skip plumbing entirely — you still own a container contract, idempotency, checkpoints and role assignments. What changes is which plumbing is yours. This session draws that line precisely, end to end, built entirely from the current Microsoft Learn documentation, with a live deployment along the way. What we'll cover • When Hosted agents are the right call — and when a prompt agent is the cheaper answer • The responsibility split — what you own inside the sandbox vs. what the platform owns around it • The three protocols — Responses, Invocations, and Invocations over WebSocket — and how to pick without regret • Sessions vs. conversations — one VM-isolated sandbox per session, scale-to-zero on idle, state restored on resume, and why these two IDs are not the same thing • Identity done right — the Entra agent identity, the project managed identity, OBO vs. application-only flows, and the role assignment that silently disappears when you publish • Toolbox in Foundry — one governed MCP endpoint for Code Interpreter, web search, Azure AI Search, OpenAPI, A2A, Skills and more, without redeploying your agent • Deploying two ways — a container image pushed to ACR, or a source zip that Foundry builds for you — plus the azd fast path, live • Versioning and rollout constraints — immutable versions, one live at a time, and no traffic splitting (so plan your canaries accordingly) • Observability and right-sizing — Application Insights wired in for free, OpenTelemetry by default, and the \~70% rule for CPU and memory • Long-running work — background execution, resilient tasks, and stream replay: what each one actually guarantees • Private networking and publishing — the VNet decision you can't undo later, and what changes when you publish to Teams and Microsoft 365