問題文
An agent repeatedly fails the same class of task. The team wants it to do better on the next attempt without retraining the model. Which approach fits that constraint?
選択肢
- Increase the maximum number of tool calls so the agent has more attempts inside a single run, because more attempts within one run is equivalent to learning from the previous run and it does not require the model to be changed.
- Have the agent write down what went wrong and carry that written reflection into the next attempt as part of its input.
- Raise the temperature so the agent tries a different approach next time.
- Fine-tune the model on the failed attempts and on a fresh set of corrected reference transcripts so the behavior is corrected.