Back to blog
Guides10 min read
A practical guide to agent observability
Traces, spans and scores: what to measure in production, which alerts matter, and how to debug an agent in minutes.
Noah BeckerHead of Engineering
Traditional monitoring tells you if a service is up. Agent observability tells you if it's doing the right thing. Here's the setup we recommend to every team.
Measure these four things
- Success rate per agent and per tool
- Latency at p50 and p95
- Cost per run
- Quality score from evals or user feedback
Alert on change, not on thresholds
Fixed thresholds are noisy. Alert when a metric moves sharply compared with the last seven days, and include a link to example traces in the alert.
“The best debugging tool is a trace you can read like a story.”