Recommending medications from electronic health records (EHRs) is a safety-critical task that demands therapeutic coverage, explicit safety constraints, traceable justification, and continual adaptivity. Supervised neural models freeze at deployment and cannot reason about new phenotypes, while existing LLM pipelines are single-pass and lack a cross-case improvement loop. We propose **SEMPA**, a training-free agentic framework that couples a Proposer with a Critic in multi-round adversarial deliberation, grounds candidates with molecular-aware retrieval, contrastively evolves both agents’ natural-language policies under a semantic-anchor constraint, and enforces rule-based DDI and contraindication checks. On 100 MIMIC-IV encounters, SEMPA reaches 0.3116 Jaccard, 0.7295 F1, and a 0.0172 DDI rate, improving F1 by 57.7% and cutting DDI by 77.5% over the strongest supervised baseline. Two LLM judges rank SEMPA highest, and six-backbone studies confirm that gains derive from the framework rather than a specific model.