Observability

What it demonstrates

An agent run is a tree of operations: the run, each model call, each tool call. Observability turns that tree into OpenTelemetry spans so you can see latency, token counts, and tool activity — printed to the console here, or exported to Jaeger / Azure Monitor in production. There is one entry point, called once at startup before any agent runs:

from agent_framework.observability import configure_otel_providers
configure_otel_providers(enable_console_exporters=True, enable_sensitive_data=True)

enable_sensitive_data=True additionally captures prompt/response/tool-arg content in the spans; setting OTEL_EXPORTER_OTLP_ENDPOINT and dropping the console flag exports to a real collector instead.

The code

async def main() -> None:
    configure_otel_providers(enable_console_exporters=True, enable_sensitive_data=True)
    agent = build_agent()
    result = await agent.run("Name three Indian classical dance forms.")
    print(f"\nAgent: {result}\n")

The console exporter then prints JSON span objects around the answer — one for the run, one for the model call.

The gotcha

Call configure_otel_providers(...) once, first, before constructing or running any agent — it wires up the global tracer providers, and later agents won’t be traced if you configure after they run. And treat enable_sensitive_data=True with care: it’s handy while learning, but it puts raw prompt and response text into your spans, which you likely don’t want leaving the process in production.

Azure / MAF mapping

The traced agent is a plain FoundryChatClient agent on Azure AI Foundry with AzureCliCredential. Observability is orthogonal to the client — the same spans emit whichever provider backs the agent, and the OTLP endpoint can point at Azure Monitor.

Run it

uv run tutorial/02-agents/04_observability.py — needs Foundry creds. It worked if JSON span objects print around the answer.

Next: Security