Conversation-Level vs Trace-Level Online Evals: Which Failures Each Can See
The unit you evaluate decides which failures you can see. A span-level evaluator judges one model call. A trace-level evaluator judges one request through the agent, including its tool calls. A conversation-level evaluator judges the whole multi-turn session after it ends. Most agent failures that…
Read the full story at DEV Community — AI ↗
Timeline · 1 report
- 2026-11-18 08:01 · DEV Community — AI
Conversation-Level vs Trace-Level Online Evals: Which Failures Each Can See