Towards autonomous medical artificial intelligence agents

The research, published in Nature, describes MIRA (Medical Intelligence for Reasoning and Action), an autonomous AI agent engineered to function within a sandboxed Electronic Health Record (EHR) system. MIRA's capabilities extend to a broad spectrum of clinical actions, including compiling patient histories, ordering and interpreting laboratory and imaging tests, formulating differential diagnoses, and devising treatment strategies like medication prescriptions and surgical scheduling. A study involving 574 simulated cases revealed MIRA achieved an impressive 88.9% diagnostic accuracy, surpassing the average 78.1% accuracy of board-certified physicians. The AI agent exhibited particular proficiency in diagnosing conditions such as appendicitis and pancreatitis, demonstrating a perfect 100% recall for specific surgical procedures. MIRA also ensured guideline-concordant and medication-safe decisions. Despite its strong performance, the study emphasizes the critical need for ongoing human oversight and further real-world clinical validation before such autonomous AI systems are fully deployed. This development signifies a major advancement in integrating AI into healthcare workflows as sophisticated decision-support partners, moving beyond basic advisory roles.