TheraAgent: Self-Improving Therapeutic Agent for Precise and Comprehensive Treatment Planning

Formulating a treatment plan is inherently a complex reasoning and refinement task rather than a simple generation problem. However, existing large language models (LLMs) mainly rely on one-shot output without explicit verification, which may result in rough, incomplete, and potentially unsafe treatment plans. To address these limitations, we propose Ther-aAgent, an agentic framework that replaces one-shot generation with an iterative generatereflect-refine pipeline. Inspired by the iterative revision patterns commonly observed in expert clinical practice, our framework progressively reflects on clinical criteria and transforms coarse drafts into precise, comprehensive and safer therapeutic regimens. To facilitate the critical reflection component, we introduce Thera-Judge, a treatment-specific evaluation module integrated into the inference loop. Experiments show TheraAgent achieves state-of-the-art results on HealthBench in treatment planning task, leading in Accuracy and Completeness. In expert evaluations, it attains an 86% win rate against physicians, with superior Targeting and Harm Control. Moreover, the high consistency between TheraJudge and HealthBench evaluation indicates the reliability of our framework.