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

CEPO. A General Framework for Improving Prompt Optimization in Multi-Agent Systems. By Jongha Kim and four others, from Korea University, KAIST, and one other institution. Published on the twenty-sixth of October.

A multi-agent system needs two things at once: effective coordination among its agents, and specialised execution of each agent's task. Existing prompt optimization approaches typically couple the two, folding both into each agent's own prompt. When one prompt carries both jobs, a rule that every agent should follow lives in many places at once, mixed in with instructions that belong to a single agent.

Coordinator-Executor Prompt Optimization, or CEPO, pulls them apart. System-level coordination goes into one coordinator prompt that every agent shares. Agent-level execution goes into executor prompts, one per agent. The updates to each kind are then guided according to its role.

The split means shared coordination rules can be refined consistently across agents, while execution instructions stay separately adjustable. CEPO is not itself a search procedure. It supplies the prompt structure and the optimization guidance, and existing prompt optimizers carry out the underlying search. That is what makes it a general framework: it sits on top of whichever optimizer is already in use.

Experiments span eleven benchmarks across four domains. Applying CEPO to four existing prompt optimizers consistently improves their performance. Testing several optimizers rather than one is what lets the gain be attributed to the structure, not to a particular search method. Ablations and qualitative analyses support the design, and controlled studies on a document-understanding multi-agent system show improved optimization budget efficiency, along with performance that scales with the number of agents.

Nothing in CEPO competes with the optimizers it wraps. Its contribution is a decision about where instructions live. Rules every agent should follow are written in one place and improved there, and the instructions that make each agent a specialist are left free to change on their own.