End-to-end request identifiers (trace/span IDs) across services so you can follow latency and errors through a call graph—not LLM-specific, but essential when AI sits in the middle of microservices.
Distributed tracing answers “which hop failed?” for multi-service flows. It complements LLM observability: traces capture infrastructure; LLM spans capture prompts, tools, and tokens when instrumented.
Standards
OpenTelemetry is the common implementation path today; pair traces with logs and chaos engineering drills.