Automated epileptic seizure detection from electroencephalography (EEG) signals is a clinically important task in which feature selection is typically performed using purely statistical criteria. We investigate whether a small instruction-tuned large language model (LLM) can guide iterative feature selection for binary seizure detection on the Epileptic Seizure Recognition dataset (11{,}500 samples, 178 features). The LLM agent (Qwen2.5-1.5B-Instruct) receives five complementary statistical summaries and selects a feature subset through multi-round reasoning. The agent achieves 96.5\% accuracy and 0.911 F1 with 40 features, compared to 97.9\% accuracy and 0.946 F1 for the best full-feature baseline (SVM-RBF on 178 features). Critically, 39 of the agent’s 40 features coincide with the top-39 mutual-information features, and a deterministic Top-39 MI filter, evaluated by the same Random Forest classifier, attains the same 96.5\% accuracy and 0.911 F1. We therefore present this work as an empirical baseline: at the 1.5B-parameter scale, the LLM behaves close to a univariate MI ranker. We situate the result against the recent LLM-based feature selection literature and enumerate the ablations and multi-dataset extensions required to determine whether larger or domain-specialized LLMs add value beyond statistical filtering.