Dictionary term
LLM Observability
Also called agent observability · LLM tracing · AI observability
Recording traces, metrics and events for every model call, tool call and agent step so you can see what an AI system did, why, and at what cost; the OpenTelemetry GenAI semantic conventions standardise the span and attribute names for this.
In plain English
A detailed activity log for your AI, showing every question it asked the model, every tool it used, how long each step took and what it cost.
Analogy
Like a flight data recorder for your agent: when something goes wrong you replay exactly what happened instead of guessing.
Why it matters
Agents fail in ways that ordinary logs miss, such as a wrong tool choice, a retry loop or a runaway token bill, so tracing is how you debug production automations, attribute cost per client and prove an agent behaved correctly.
In practice
Exporting Claude Code or agent telemetry over OpenTelemetry to your monitoring stack, then filtering traces to find the run where a CRM-update tool was called twice for the same contact.
Related terms
Sources
Browse all terms →Last verified