AI Agents and Multi-Agent Systems · Lesson 11
Agent Evaluation & Debugging
Evaluate agents on outcomes (did the task succeed?), trajectories (were the steps and tool calls sensible?), and efficiency (steps, tokens, time, cost).
Tracing tools such as LangSmith, Langfuse or Arize Phoenix record every model call and tool call so you can replay a run and see where it went wrong.
Common failures: wrong tool choice, bad arguments, infinite loops, lost context and premature stopping. Fix with better tool descriptions, clearer prompts, step limits and targeted tests.
Key points
- Measure outcome, trajectory and cost
- Trace every model and tool call
- Replay failures to find the broken step
- Turn bugs into regression tests