Generative multi-agent systems are emerging as a powerful paradigm for simulating human-like behavior in real-time applications such as interactive storytelling, virtual reality environments, and autonomous decision-making. These agents, often powered by large language models and memory systems, act independently and adapt over time. However, a critical challenge in deploying such systems is ensuring their fault tolerance. The ability to maintain operation in the presence of faults such as communication failures, memory corruption, agent crashes, or behavioral inconsistencies. This paper presents a comprehensive review of fault tolerance techniques for generative agents, focusing on methods such as memory check pointing, agent replication, fusion-based resilience, and consistency protocols. We analyse these approaches, drawing parallels from distributed systems, and evaluate their effectiveness in maintaining operational integrity in large-scale, real-time environments. Our findings suggest that while no single technique offers a one-size-fits-all solution, a combination of methods can provide robust fault tolerance and support the scalability and reliability of generative agent systems in dynamic, fault-prone environments.
Uddin et al. (Thu,) studied this question.