The paper introduces Agent Hospital, a simulacrum of a hospital environment where patients, nurses, and doctors are autonomous agents powered by large language models (LLMs). The goal is to enable doctor agents to learn and improve their medical skills within this simulated environment. The proposed method, MedAgent-Zero, allows doctor agents to evolve without manually labeled data, accumulating experience from both successful and unsuccessful cases. Experiments show that the doctor agents' performance consistently improves on various tasks, achieving state-of-the-art accuracy on a subset of the MedQA dataset for respiratory diseases after treating around ten thousand simulated patients. This work paves the way for advancing LLM-powered agent techniques in medical scenarios.The paper introduces Agent Hospital, a simulacrum of a hospital environment where patients, nurses, and doctors are autonomous agents powered by large language models (LLMs). The goal is to enable doctor agents to learn and improve their medical skills within this simulated environment. The proposed method, MedAgent-Zero, allows doctor agents to evolve without manually labeled data, accumulating experience from both successful and unsuccessful cases. Experiments show that the doctor agents' performance consistently improves on various tasks, achieving state-of-the-art accuracy on a subset of the MedQA dataset for respiratory diseases after treating around ten thousand simulated patients. This work paves the way for advancing LLM-powered agent techniques in medical scenarios.