OpenTelemetry has two competing semantic conventions for tracing agent behavior: the OTEL GenAI standard and OpenInference. We've found OpenInference to be significantly more useful for production agent workloads, offering richer LLM call metadata, first-class support for RAG and retrieval spans, and better distinctions between span types. We've found that projects that use OpenInference vs. OTEL GenAI standard are better positioned to improve individual components of their agentic architecture (eg: Prompts, Tool Calls, RAG, etc.), allowing them to ship faster and with more confidence. This post covers why, along with what to instrument and which frameworks give you strong observability out of the box.

Interesting