Community GitHub guide details how to host agents on Microsoft Foundry

Microsoft Foundry’s hosted-agent story is getting a lot more concrete. The big change is not some flashy model launch. It’s that Microsoft’s docs and GitHub ecosystem now give developers a pretty clear path from local agent code to a running service inside Foundry. That closes a real gap. Plenty of teams liked the idea of “agents,” but the hard part was always deployment — packaging code, wiring identity, handling scaling, and keeping the whole thing observable once it went live. ### What is a hosted agent, exactly? A hosted agent in Foundry is not just a prompt with a tool attached. It’s a containerized application that runs your own code on Microsoft-managed infrastructure, while Foundry handles the surrounding platform pieces — scaling, lifecycle, session state, and observability. Your code does the orchestration. Foundry provides the runtime, endpoint, and managed access to models and tools. ### Why does that matter more than a prompt agent? Prompt agents are the quick path. They work when the logic mostly lives in instructions and built-in tools. Hosted agents are for the harder cases — custom frameworks, webhooks, stateful workflows, nonstandard payloads, or apps that need specific CPU and memory settings. Basically, if your agent behaves more like software than a chat preset, hosted agents are the fit. ### What changed in practice? The practical shift is that Microsoft now has current quickstarts, deployment docs, and public GitHub repos that walk through the process end to end. The quickstart covers initializing a hosted-agent project with the Azure Developer CLI or VS Code. The GitHub side now includes a Foundry Toolkit workshop repo and a.NET demo repo with staged scenarios, from theoretical and much more buildable. ### How does deployment actually work? The flow is pretty straightforward. You package the agent as a container image, push it to Azure Container Registry, register that image with Foundry Agent Service, wait for the deployment to become active, and then call the dedicated endpoint. If you use the quickstart path, azd or the VS Code extension handles a lot of the annoying plumbing — image build, push, versioning, and much of the role setup. ### Where does security come in? Security is one of the main reasons this matters. Each hosted agent gets a dedicated Microsoft Entra identity at deploy time, and that identity is what the running container uses to call models and tools. Foundry also layers in RBAC, managed authentication, content filters, and optional virtual network isolation. The catch is that per-deployment created identity. ### What about observability? This is the part teams usually underestimate. Foundry says Agent Service includes end-to-end tracing, metrics, and Application Insights integration. In plain English, that means you’re not just deploying an agent and hoping for the best. You can inspect decisions, trace tool calls, and monitor behavior in production — which is the difference between a demo and an operable system. ### Is multi-cloud part of the story? Sort of — but as an inference, not a formal product announcement. Foundry’s hosted-agent docs explicitly frame the problem around “heterogeneous cloud environments,” and the platform is built to let your code orchestrate external services while Foundry handles hosting and identity. That means teams can use Foundry as the managed runtime even if they, though