This paper explores the challenges and open problems in multi-agent systems (MAS), emphasizing their potential to enhance the capabilities of single LLM agents through collaboration. MAS leverage the diverse roles and specializations of individual agents to tackle complex tasks. Key challenges include optimizing task allocation, fostering robust reasoning through iterative debates, managing complex context information, and improving memory management. The paper also discusses the potential applications of MAS in blockchain systems, highlighting their ability to support distributed computing environments. MAS can be structured in various ways, including equi-level, hierarchical, nested, and dynamic structures. Each structure presents unique challenges in terms of coordination, communication, and memory management. For instance, hierarchical structures involve a leader and followers, while dynamic structures allow agents to change roles and interactions over time. Planning in MAS involves global planning, where tasks are divided and coordinated among agents, and local planning, where individual agents break down tasks into smaller steps. Challenges include aligning goals, managing complex contexts, and ensuring consistency across agents. Game theory provides a framework for strategic interactions, particularly in scenarios involving debates or discussions. Memory management in MAS is more complex than in single-agent systems, requiring sophisticated mechanisms for sharing, integrating, and managing information. Different types of memory, such as short-term, long-term, and consensus memory, are crucial for effective collaboration. Challenges include ensuring data security, maintaining consensus, and managing information exchange between agents. Applications of MAS in blockchain systems are explored, including smart contract analysis, consensus mechanism enhancement, and fraud detection. By assigning agents to blockchain nodes, MAS can enhance data processing, security, and privacy. The paper concludes that MAS have significant potential to advance LLM capabilities and improve distributed computing environments, but further research is needed to address the challenges in planning, memory management, and system integration.This paper explores the challenges and open problems in multi-agent systems (MAS), emphasizing their potential to enhance the capabilities of single LLM agents through collaboration. MAS leverage the diverse roles and specializations of individual agents to tackle complex tasks. Key challenges include optimizing task allocation, fostering robust reasoning through iterative debates, managing complex context information, and improving memory management. The paper also discusses the potential applications of MAS in blockchain systems, highlighting their ability to support distributed computing environments. MAS can be structured in various ways, including equi-level, hierarchical, nested, and dynamic structures. Each structure presents unique challenges in terms of coordination, communication, and memory management. For instance, hierarchical structures involve a leader and followers, while dynamic structures allow agents to change roles and interactions over time. Planning in MAS involves global planning, where tasks are divided and coordinated among agents, and local planning, where individual agents break down tasks into smaller steps. Challenges include aligning goals, managing complex contexts, and ensuring consistency across agents. Game theory provides a framework for strategic interactions, particularly in scenarios involving debates or discussions. Memory management in MAS is more complex than in single-agent systems, requiring sophisticated mechanisms for sharing, integrating, and managing information. Different types of memory, such as short-term, long-term, and consensus memory, are crucial for effective collaboration. Challenges include ensuring data security, maintaining consensus, and managing information exchange between agents. Applications of MAS in blockchain systems are explored, including smart contract analysis, consensus mechanism enhancement, and fraud detection. By assigning agents to blockchain nodes, MAS can enhance data processing, security, and privacy. The paper concludes that MAS have significant potential to advance LLM capabilities and improve distributed computing environments, but further research is needed to address the challenges in planning, memory management, and system integration.