Agent Observability

Overview

Production agents fail differently than services: bad tool calls, hallucinated arguments, runaway loops, silent quality regression. This skill picks an observability backend, instruments node-level spans, attaches evals to traces, and sets up replay — so failures are observable, not just guessed at.

When to use

• Agent works in dev, breaks in prod with no logs that explain why
• Need to compare prompt/model changes against a baseline (eval-in-trace)
• Token/latency cost is growing and you don't know which node is the culprit
• A user reports a bad answer and you need to replay the exact session
• Multiple agents/subagents — you need a single trace tree, not interleaved logs

Platform selection