# Multi-Agent Design: Optimizing Agents with Better Prompts and Topologies

**Authors:** Han Zhou, Xingchen Wan, Ruoxi Sun, Hamid Palangi, Shariq Iqbal, Ivan Vulić, Anna Korhonen, Sercan Ö. Arık  
**arXiv:** [2502.02533](https://arxiv.org/abs/2502.02533)  
**Date:** 2025-02-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2502.02533v1/

## Key points

- Large language models function as multiple agents collaborating on tasks
- Agents are equipped with prompts defining functions and topologies facilitating interactions
- Designing prompts and topologies for multi-agent systems (MAS) requires an automated process
- Authors analyze key factors influencing effective MAS construction
- Prompts and topologies play pivotal roles in enhancing MAS design efficacy
- Introduction of Multi-Agent System Search (MASS) optimization framework
- MASS optimizes MAS through block-level prompt optimization, workflow topology optimization, and workflow-level prompt optimization stages
- MASS-optimized multi-agent systems outperform existing alternatives significantly
- Proposal of fundamental design principles for constructing effective multi-agent systems based on MASS findings
- Study contributes valuable knowledge to the field of multi-agent systems research

## Summary

In their paper titled "Multi-Agent Design: Optimizing Agents with Better Prompts and Topologies," authors Han Zhou, Xingchen Wan, Ruoxi Sun, Hamid Palangi, Shariq Iqbal, Ivan Vulić, Anna Korhonen, and Sercan Ö. Arık delve into the realm of large language models functioning as multiple agents collaborating to tackle intricate tasks. These agents are equipped with prompts that define their functions and topologies that facilitate interactions among them. The complexity inherent in designing prompts and topologies for multi-agent systems (MAS) necessitates an automated design process. To address this challenge, the authors conduct a thorough analysis of the design space to comprehend the key factors influencing effective MAS construction. Their investigation uncovers the pivotal roles played by prompts and topologies in enhancing MAS design efficacy. Building on these insights, they introduce Multi-Agent System Search (MASS), a MAS optimization framework designed to efficiently navigate the intricate MAS design space. MASS optimizes MAS through three stages: 1) block-level prompt optimization at a local level; 2) workflow topology optimization; and 3) workflow-level prompt optimization at a global scale. Each stage builds upon the iteratively optimized prompts and topologies from preceding stages. The authors demonstrate that MASS-optimized multi-agent systems outperform existing alternatives by a significant margin. Drawing from the systems identified through MASS optimization, they propose fundamental design principles for constructing effective multi-agent systems. This comprehensive approach not only advances our understanding of MAS design but also offers practical insights for optimizing agent performance through refined prompts and topologies. The study's findings contribute valuable knowledge to the field of multi-agent systems research, paving the way for enhanced efficiency and effectiveness in complex task-solving scenarios.