LLM Multi-Agent Systems: Challenges and Open Problems

The paper provides an extensive overview of Large Language Model (LLM) Multi-Agent Systems, exploring their potential and the significant challenges that remain inadequately addressed compared to single-agent systems. The authors examine key issues such as optimising task allocation to leverage agents' specialised skills, fostering robust reasoning through iterative debate, and effectively managing complex, layered context information and various forms of memory. Furthermore, the paper categorises the structural configurations of these systems— including equi-level, hierarchical, nested, and dynamic—and discusses potential real-world applications in blockchain systems, such as smart contract analysis and consensus mechanism enhancement. This work ultimately surveys the current landscape of multi-agent systems while highlighting critical areas for future research, including global and local planning, and advanced memory management.