A new academic framework called EAGLET proposes a practical and efficient method to improve long-horizon task performance in LLM-based agents — without the need for manual data labeling or retraining.