Run on AKS. Trace in Foundry.¶

One Pydantic AI agent, an authenticated HTTP endpoint, and one model call per request. Foundry adds an external-agent record and attributed traces. Execution stays on AKS.

Microsoft public preview

This repository uses Microsoft's external-agent registration for observability and evaluation (preview). Foundry stores registration metadata and matches OpenTelemetry traces by gen_ai.agent.id; it does not host, proxy, or invoke the AKS runtime.

Microsoft provides this preview without a service-level agreement and does not recommend it for production workloads. This repository implements registration, tracing, synthetic traffic generation, and one-off trace evaluation plus on-demand Insights.

The CLI reaches an authenticated service and Pydantic AI agent inside AKS through a loopback port-forward. The agent calls the Foundry model. Separately, agent telemetry is exported to project-linked Application Insights and viewed in Foundry.

Open diagram full-size Architecture and sources

Understand the design¶

Follow the request and telemetry paths, then review the authentication and security boundaries. Path A is metadata and observability integration, not another runtime.

Deploy and verify¶

Set up local development, provision the dedicated infrastructure, then deploy and verify an authenticated model request.

What verification covers¶

The verification guide connects an authenticated model answer to its image digest, source release, pod, stable agent ID, and native request/agent/model trace chain. It requires the same trace in the connected Application Insights resource and the external agent's Foundry trace view.

Deliberately small¶

This is a single-operator, synthetic-data reference deployment. It has no agent tools, conversation store, retries, or public application ingress. Path B (APIM/AI Gateway), scheduled Insights or evaluation, multi-turn evaluation, document search, ingestion, and an agent web UI are outside this repository. See the runtime boundaries.