This paper explores the integration of Large Language Models (LLMs) with human agents to enhance the solving of complex tasks. The authors address the limitations of LLM-based agents in dynamic environments and propose a Reinforcement Learning-based Human-Agent Collaboration method, ReHAC, to improve task-solving performance through well-planned human intervention. ReHAC is trained using a dataset of human-agent collaborations and is evaluated on three multi-step reasoning datasets: HotpotQA, StrategyQA, and InterCode. The results show that ReHAC effectively balances task performance and human intervention, outperforming other baselines such as agent-only, human-only, random, prompt, and imitation learning methods. The paper also discusses the potential challenges and future directions, including multi-level human-agent collaboration, the development stages of LLM-based agents, and safety and super alignment.This paper explores the integration of Large Language Models (LLMs) with human agents to enhance the solving of complex tasks. The authors address the limitations of LLM-based agents in dynamic environments and propose a Reinforcement Learning-based Human-Agent Collaboration method, ReHAC, to improve task-solving performance through well-planned human intervention. ReHAC is trained using a dataset of human-agent collaborations and is evaluated on three multi-step reasoning datasets: HotpotQA, StrategyQA, and InterCode. The results show that ReHAC effectively balances task performance and human intervention, outperforming other baselines such as agent-only, human-only, random, prompt, and imitation learning methods. The paper also discusses the potential challenges and future directions, including multi-level human-agent collaboration, the development stages of LLM-based agents, and safety and super alignment.