Engineers praise Microsoft Foundry Agents public preview for hosted, secure agent execution

Microsoft is trying to make AI agents feel less like a science project and more like normal cloud software. That is the point of the refreshed public preview for hosted agents in Foundry Agent. Try one of these systems yourself. The model is the easy part. The ugly part is packaging code, locking it down, scaling it, tracing it, and keeping out of that mess. Hosted agents in Foundry Agent Service let developers deploy containerized agents onto Microsoft-managed infrastructure instead of standing up their own runtime stack. The service handles hosting, deployment, scaling, and observability, while the agent calls models from the Foundry model catalog and runs custom orchestration code around them. Microsoft frames this as a production path for enterprise agents, not just a playground feature. ### Why are engineers reacting to the “hosted” part? Because self-hosting agents is where the pain usually starts. Teams want to absorb those cross-cutting concerns into the platform. In plain English, teams can spend more time on tools, prompts, and workflows, and less time babysitting Kubernetes-shaped problems. ### What changed in this preview? This is not just a minor rename. Microsoft says the refreshed preview uses a new hosting backend, new protocol libraries, a new identity model, and new migration tooling; existing deployments do not roll forward automatically. Teams have to redeploy onto the new model. ### Why does the security angle matter? The standout feature is isolated execution per session. Microsoft says each agent session gets its own dedicated sandbox with a persistent file system — not just process isolation. That matters for enterprise use because when an agent runs for a company, you need strong boundaries around every run, not a shared runtime with fingers crossed. ### What does Microsoft think the cost story is? The company is pitching “scale-to-zero” economics, which is cloud workloads — internal copilots, workflow agents, or task runners that sit quiet until someone invokes them. Managed scaling is not glamorous, but it is the kind of feature that decides whether a prototype survives budget review. ### Is there real developer scaffolding yet? Yes — and that is part of why the samples exist. The samples come with Application Insights and OpenTelemetry tracing, and the quickstart points developers to the Foundry Toolkit for VS Code for testing and interaction. This is not just a landing page and a promise. ### What is the catch? Preview means churn. Microsoft’s migration guide is blunt that the old backend will only be supported until May 22, 2026, and there is no question Microsoft’s position is strong, but the operational contract is not fully settled yet. ### Bottom line? The excitement here is not that Microsoft invented agents. It is that Microsoft is packaging the boring, hard, enterprise parts — isolation, identity, deployment, scaling, tracing — into a service that looks a lot closer to normal infrastructure. If that holds up, this becomes less a demo feature and more the default way big companies run agent workloads on Azure.