Problem
The wrong answer comes back with a 200 status. Logs show latency and tokens, but not what went into the model or why.
Solution
Each interaction produces a trace with a span per step, prompt and model versions and masked context. The trace feeds evaluation and cost attribution.
When to use
- →As soon as there are real users
- →Multi-step agents
- →Need to explain decisions
## Trade-offs
- +Reproducible incidents
- +Cost per task and per customer
- +Foundation for continuous evaluation
- −Stores sensitive data without masking
- −Data volume requires sampling