A practical guide to setting up distributed tracing for multi-agent AI workflows using OpenTelemetry, based on lessons from building the it-self-service-agent AI quickstart. Covers context propagation across service boundaries using W3C Trace Context, auto-instrumentation for FastAPI and HTTPX clients, manual instrumentation for MCP servers using a decorator pattern, Llama Stack telemetry configuration for versions 0.2.x and 0.3.x, and deployment options ranging from Jaeger All-in-One for development to Red Hat OpenShift Distributed Tracing for production. Includes concrete code examples for span creation, context extraction/injection, error handling, and attribute best practices.
Table of contents
About AI quickstartsDistributed tracing for agentic workloadsWhat is OpenTelemetry?Context propagationInstrumenting Llama StackAuto-instrumentation: HTTP clients and FastAPIManual instrumentation: MCP serversLlama Stack tracing configurationCollect traces with a Jaeger All-in-One serverCollect traces with Red Hat OpenShift Distributed TracingWrapping upSort: