GLOSSARY
Observability
Seeing inside a running LLM app — traces, costs, failures — so you can debug agents the way you debug services.
Classical observability is logs, metrics and traces. LLM observability adds the payload: the prompt that went in, the tool calls in between, the tokens and dollars that came out, and whether the answer was any good. Without that loop, production agents fail silently three steps before the user-visible error.
LangSmith, Langfuse and Helicone are the usual layers. The point is not another dashboard — it is making prompt changes testable and regressions visible. Teams that skip it ship vibes; teams that keep traces can say why last Tuesday's answers got worse.