CEPO splits multi-agent prompt optimization into a shared coordinator prompt and per-agent executor prompts, and it's built to sit on top of existing optimizers.
The split: Existing approaches couple coordination and task execution inside agent-specific prompts. CEPO keeps system-level coordination in one shared coordinator prompt and execution instructions in agent-specific executor prompts, so shared rules can be refined across agents.
Role: CEPO supplies the prompt structure and optimization guidance, while existing prompt optimizers do the underlying search.
Results: The paper reports experiments across 11 benchmarks in four domains, where applying CEPO to four existing optimizers improved their performance.
Efficiency: On a document-understanding MAS, the paper's controlled studies show better optimization budget efficiency and performance that scales with the number of agents.
Sign in to comment.
TL;DR of the paper, generated automatically.
CEPO splits multi-agent prompt optimization into a shared coordinator prompt and per-agent executor prompts, and it's built to sit on top of existing optimizers.