The integration of AI-powered agents in medical education offers a significant opportunity to enhance communication skills among future physicians. This study introduces a novel dual-agent AI system, leveraging large language models, to simulate realistic patient interactions during a neurology visit for memory loss and to evaluate communication proficiency. The patient simulator engages medical students in history-taking exercises, while the evaluator agent provides immediate, actionable insights and structured feedback directly aligned with the Core Entrustable Professional Activities (EPAs) framework. Findings from a participatory design process with three neuroscience experts indicate that while the agent effectively delivers medical information and fosters dialogue, limitations remain in maintaining nuanced contextual understanding and conveying empathetic responses. We discuss these limitations and propose refinements to the model and feedback mechanisms to enhance its educational efficacy. This research underscores the potential of dual AI agents in teaching crucial doctor-patient communication skills, contributing to the advancement of AI-assisted personalized medical education, and highlighting key areas for future development.