I built agents with the LangSmith SDK and in Fleet. Tracing exposed a retry loop and a tool-input bug; a prompt comparison scored 1.00 on both variants while one cost three times as much; and results got worse after a skill revision, which is why every skill version needs its own evaluation.