CEPO: General Framework for Improving Prompt Optimization in Multi-Agent Systems (arxiv.org)

The recent introduction of the Coordinator-Executor Prompt Optimization (CEPO) framework marks a significant advancement in the field of multi-agent systems (MAS), optimizing prompt strategies for improved coordination and task execution. Unlike traditional methods that fuse these functions within agent-specific prompts, CEPO innovatively separates system-level coordination from agent-level execution. This approach utilizes a shared coordinator prompt to guide updates while allowing for the customization of each agent's execution prompts, enhancing both efficiency and adaptability. The implications of CEPO are profound for the AI/ML community, as demonstrated by extensive experiments across 11 benchmarks in four domains, revealing a consistent performance boost when integrating CEPO with existing prompt optimizers. Notably, the framework improves optimization budget efficiency and scales performance with the number of agents, which could drive significant advancements in the deployment of MAS in complex, real-world tasks. This separation of roles and guided updates not only refines collaborative capabilities but also sets a new standard for prompt optimization in AI systems, making CEPO a pivotal development for future research and applications in multi-agent environments.