Skills
Evaluate agents at three levels: end-to-end, trajectory, and component
Single 'did the task succeed?' scores hide where an agent actually broke — 2026 practice evaluates at end-to-end (task success), trajectory (was the path efficient and sound?), and component level (which retriever, tool, or subagent failed). Trace-level visibility over every LLM call, tool invocation, and retrieval step is the backbone that surfaces new failure modes you don't yet have metrics for. Instrument all three so a regression points you at the exact broken component, not just a lower aggregate.
Source
↳ Follow the thread