Foundry Agent Service reached general availability in July 2026, and the best part of the story is that you do not have to rewrite your agent for it. In this session we take an existing agent written in LangGraph and host it as a Foundry hosted agent without touching its logic: every session runs in its own VM-isolated sandbox, and the agent automatically receives its own Entra identity, a dedicated endpoint, a persistent home directory, and OpenTelemetry tracing. Then we go where the marketing ends: through the full azd lifecycle from init to deploy, through generating and running evaluations, to the agent optimizer, which proposes improvements as ranked candidates with a complete audit trail and one-command rollback. Along the way we tick off the practical traps: immutable versions with no traffic splitting, sandbox size ceilings, billing by active sessions, private container registry rules, choosing between the Responses and Invocations protocols, and how all of this compares to Container Apps and AKS. The talk is grounded in a real architecture decision from a production project and is aimed at everyone moving agents from prototypes into governed, managed production on Azure.