Tracing, logging, and metrics tailored to model calls: latency, token usage, tool errors, retrieval quality, and evaluation signals.
Classic APM misses prompt/response structure. LLM observability captures spans per call, tool round-trips, and retrieval hits so you can debug failures and regressions.
Why teams adopt it
When agents and RAG enter production, “it worked in the demo” is not enough—you need repeatable forensics and dashboards tied to releases.