CEPO: A General Framework for Improving Prompt Optimization in Multi-Agent Systems

Jongha Kim^1,*   Seohyun Lee^2,*   Jonathan Minho Chung²   Jinsung Yoon³   Hyunwoo J. Kim^2,†

¹ Korea University    ² KAIST    ³ Google Cloud AI

_{* Equal contribution    † Corresponding author}

_{Teaser (1:40). 1080p mp4 · Project page}

CEPO splits the prompts of a multi-agent system into a shared coordinator prompt (system-level coordination) and per-agent executor prompts (agent-level execution), and optimizes the two layers sequentially with role-aligned guidance on top of an existing prompt optimizer (PE2, OPRO, ProTeGi, ...).

Code

🚧 Coming soon.

Citation

@article{kim2026cepo,
  title   = {CEPO: A General Framework for Improving Prompt Optimization in Multi-Agent Systems},
  author  = {Kim, Jongha and Lee, Seohyun and Chung, Jonathan Minho and Yoon, Jinsung and Kim, Hyunwoo J.},
  journal = {arXiv preprint arXiv:2610.34722},
  year    = {2026}
}