Observability
Last verified
Observability for LLM applications means traces, spans, prompt/response logging and evaluation. It answers _why did the model do that_.
Worth separating from cost tracking, because the two are sold together and are different jobs. If your question is "what are we spending", billing APIs answer it with no SDK and no proxy. If your question is "why did my agent loop", you need traces.
A useful test: in the last month, did you open a trace? If not, you want a spend number. Langfuse alternatives.