Skip to content
anythingai.space
Frameworks & Toolsintermediate11 min · 9 steps

Observability with LangSmith

Trace every node of an agent run, turn feedback into datasets, score runs with judges, compare variants, and budget for tracing itself.

WHAT THE LOGS SAYcalling modeldonecalling modeldone200 OKUserResearcherwrong URL?SummarizerhallucinatesAnswer
requestdataerrorsignalresponse
1 / 9
ProblemStep 1 of 9

The answer is wrong. Which step broke?

A user reports a confidently wrong answer. The pipeline has a researcher, a summarizer and a final writer, and the logs say little more than calling model… done.

The real error happened many steps earlier: the researcher fetched the wrong URL, so the summarizer had nothing true to work with. Without the full trajectory, you end up tweaking the wrong prompt.

Lesson 4 of 6 in Frameworks & Tools