<p>Industrial classification tasks often face challenges such as class imbalance, noise and non-stationary data distributions. Most feature engineering methods aided by evolutionary algorithms and large language models (LLMs) of-ten rely on the predictive performance of downstream classification metrics, while neglecting the feature distribution structure and the relationship between features and labels under distribution shifts. To address these issues, we propose the Feature Meta-Model Agent (FMM-Agent), a framework that evolves features within a meta-model space defined by statistical information, rather than operating directly on raw data. FMM-Agent enables LLMs to perform operator restructure and refine the chain-of-thought to obtain better feature shaping. We further introduce a unified scoring mechanism to jointly evaluate label relevance and distribution stability, allowing the feature pool to gradually move toward better feature distribution shapes. Experiments conducted on 11 data sets show that FMM-Agent consistently improves the recognition ability of minority classes in terms of balanced accuracy, g-mean, and recall, and its performance is superior to other comparative methods. Ablation studies confirm the necessity of evolutionary restructure and generation mechanism strategies. In addition, experimental results with different LLMs show that although stronger models can produce more stable evolutionary process, the overall performance improvement of FMM-Agent does not depend on a specific model. It is worth noting that although FMM-Agent incurs additional inference time costs due to evolutionary feature generation, it achieves a good balance be-tween computational overhead and performance improvement.</p>