Task success, tool use, steps, cost — and traces

Run the 30 tasks. Score: correct final answer (compare numbers with your expected answers; LLM-judge for wording), correct tools used, approval requested when needed, refusals correct, steps and cost per task. Turn on tracing (LangSmith or Langfuse) and read the traces of every failed task: wrong tool, wrong arguments, gave up, looped, made up a number. Fix the cause (tool description, prompt, tool output) and re-run.