Agent-Pro is an LLM-based agent that learns and evolves through policy-level reflection and optimization. It can learn from interactive experiences and improve its behavioral policy over time. The agent uses dynamic belief generation and reflection to iteratively refine its irrational beliefs and optimize its policy. A depth-first search is employed for policy optimization, ensuring continuous improvement in policy payoffs. Agent-Pro is evaluated in two games: Blackjack and Texas Hold'em, outperforming vanilla LLMs and specialized models. The results show that Agent-Pro can learn and evolve in complex and dynamic scenarios, which benefits many LLM-based applications. The agent is designed to handle multi-player imperfect information interactive games, where agents have limited information about the environment and other players. Agent-Pro uses a belief-aware decision-making process, which involves updating beliefs about the world and itself, and then making decisions based on these beliefs. It also employs policy-level reflection to correct irrational beliefs and optimize its policy. The agent's policy is refined through a depth-first search, allowing it to incrementally improve its performance. In Blackjack, Agent-Pro significantly outperforms baseline agents, showing improved decision-making capabilities. In Limit Texas Hold'em, Agent-Pro surpasses other LLM-based agents and even RL-based agents, demonstrating its ability to learn and adapt in complex scenarios. The agent's performance is enhanced by its dynamic belief, which allows it to adjust to changes in the environment and opponents' strategies. Agent-Pro's learning process involves policy-level reflection and optimization, which enable it to improve its performance over time. The agent's policy is refined through a depth-first search, allowing it to find better strategies. The agent's ability to learn and evolve makes it suitable for complex tasks beyond card games, such as business, company negotiations, and security. Despite its success, Agent-Pro has limitations, including its reliance on the foundational model's capabilities and potential gaps compared to state-of-the-art algorithms. However, the agent's ability to learn and evolve through policy-level reflection and optimization makes it a promising approach for LLM-based agents in complex tasks.Agent-Pro is an LLM-based agent that learns and evolves through policy-level reflection and optimization. It can learn from interactive experiences and improve its behavioral policy over time. The agent uses dynamic belief generation and reflection to iteratively refine its irrational beliefs and optimize its policy. A depth-first search is employed for policy optimization, ensuring continuous improvement in policy payoffs. Agent-Pro is evaluated in two games: Blackjack and Texas Hold'em, outperforming vanilla LLMs and specialized models. The results show that Agent-Pro can learn and evolve in complex and dynamic scenarios, which benefits many LLM-based applications. The agent is designed to handle multi-player imperfect information interactive games, where agents have limited information about the environment and other players. Agent-Pro uses a belief-aware decision-making process, which involves updating beliefs about the world and itself, and then making decisions based on these beliefs. It also employs policy-level reflection to correct irrational beliefs and optimize its policy. The agent's policy is refined through a depth-first search, allowing it to incrementally improve its performance. In Blackjack, Agent-Pro significantly outperforms baseline agents, showing improved decision-making capabilities. In Limit Texas Hold'em, Agent-Pro surpasses other LLM-based agents and even RL-based agents, demonstrating its ability to learn and adapt in complex scenarios. The agent's performance is enhanced by its dynamic belief, which allows it to adjust to changes in the environment and opponents' strategies. Agent-Pro's learning process involves policy-level reflection and optimization, which enable it to improve its performance over time. The agent's policy is refined through a depth-first search, allowing it to find better strategies. The agent's ability to learn and evolve makes it suitable for complex tasks beyond card games, such as business, company negotiations, and security. Despite its success, Agent-Pro has limitations, including its reliance on the foundational model's capabilities and potential gaps compared to state-of-the-art algorithms. However, the agent's ability to learn and evolve through policy-level reflection and optimization makes it a promising approach for LLM-based agents in complex tasks.