This paper introduces a novel automated feature generation method called LLM Feature Generation (LFG), which leverages large language models (LLMs) to enhance the interpretability, adaptability, and strategic flexibility of automated feature engineering. The method addresses three key challenges in traditional feature generation: lack of explainability, limited applicability, and rigid strategy formulation. LFG employs a dynamic and adaptive approach, using LLMs with expert-level guidance to generate features that are both interpretable and effective for downstream tasks. The method involves creating multiple expert agents that generate new features through various operations, and iteratively refining these features based on performance feedback from downstream tasks. The agents continuously learn and adjust their strategies to optimize the feature space, leading to improved performance in machine learning tasks. The method is evaluated on multiple real-world datasets and demonstrates significant improvements over existing feature generation techniques. The results show that LFG outperforms baseline methods in terms of accuracy, precision, recall, and F1 scores across various classification tasks. The method also exhibits robustness and adaptability, with performance improvements continuing through multiple iterations. The paper concludes that LFG provides a promising approach to automated feature engineering, with potential for broader application in diverse machine learning scenarios.This paper introduces a novel automated feature generation method called LLM Feature Generation (LFG), which leverages large language models (LLMs) to enhance the interpretability, adaptability, and strategic flexibility of automated feature engineering. The method addresses three key challenges in traditional feature generation: lack of explainability, limited applicability, and rigid strategy formulation. LFG employs a dynamic and adaptive approach, using LLMs with expert-level guidance to generate features that are both interpretable and effective for downstream tasks. The method involves creating multiple expert agents that generate new features through various operations, and iteratively refining these features based on performance feedback from downstream tasks. The agents continuously learn and adjust their strategies to optimize the feature space, leading to improved performance in machine learning tasks. The method is evaluated on multiple real-world datasets and demonstrates significant improvements over existing feature generation techniques. The results show that LFG outperforms baseline methods in terms of accuracy, precision, recall, and F1 scores across various classification tasks. The method also exhibits robustness and adaptability, with performance improvements continuing through multiple iterations. The paper concludes that LFG provides a promising approach to automated feature engineering, with potential for broader application in diverse machine learning scenarios.