CEPO (Coordinator‑Executor Prompt Optimization) introduces a unified framework that separates system‑level coordination from agent‑level execution in multi‑agent systems. The framework consists of a shared coordinator prompt that encodes global collaboration rules and individual executor prompts that contain task‑specific instructions for each agent. Updates are guided by the distinct roles: the coordinator prompt is refined consistently across agents, while executor prompts are adjusted independently to match each agent's capabilities.
CEPO does not replace existing prompt optimizers; it supplies the prompt architecture and role‑aware guidance, leaving the underlying search to established optimizers. Applying CEPO to four state‑of‑the‑art optimizers across 11 benchmarks in four domains consistently improves their results. Ablation studies confirm that shared coordination and separate execution tuning are the primary contributors to the gains.
Controlled experiments on a document‑understanding multi‑agent setup demonstrate that CEPO reduces the required optimization budget and scales performance roughly linearly with the number of agents.
Review