# MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework

**Authors:** Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Ceyao Zhang, Jinlin Wang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, Chenyu Ran, Lingfeng Xiao, Chenglin Wu, Jürgen Schmidhuber  
**arXiv:** [2308.00352](https://arxiv.org/abs/2308.00352)  
**Date:** 2023-08-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2308.00352v5/

## Key points

- Significant advancements in automated problem-solving through societies of agents based on Large Language Models (LLMs)
- Challenges faced by existing LLM-based multi-agent systems in tackling complex tasks due to logic inconsistencies from cascading hallucinations
- Introduction of MetaGPT, a meta-programming framework incorporating Standardized Operating Procedures (SOPs) for streamlined workflows
- MetaGPT's adoption of an assembly line paradigm assigning diverse roles to agents for breaking down complex tasks into subtasks
- Demonstration of MetaGPT's ability to generate more coherent solutions compared to previous chat-based multi-agent systems on collaborative software engineering benchmarks
- Emphasis on improving task decomposition and coordination among agents through SOPs and structured communication protocols within MetaGPT

## Summary

, , , , 

Significant advancements have been made in the realm of automated problem-solving through societies of agents based on Large Language Models (LLMs). While existing LLM-based multi-agent systems have shown success in solving simple dialogue tasks, tackling more complex tasks has proven challenging due to logic inconsistencies caused by cascading hallucinations from naively chaining LLMs. To address this issue, MetaGPT, an innovative meta-programming framework, has been introduced. MetaGPT incorporates efficient human workflows into LLM-based multi-agent collaborations by encoding Standardized Operating Procedures (SOPs) into prompt sequences for streamlined workflows. This allows agents with human-like domain expertise to verify intermediate results and reduce errors. MetaGPT adopts an assembly line paradigm that assigns diverse roles to various agents, breaking down complex tasks into subtasks involving multiple agents working together. On collaborative software engineering benchmarks, MetaGPT has demonstrated the ability to generate more coherent solutions compared to previous chat-based multi-agent systems. The project is led by a team of researchers and developers who have contributed significantly to its design and implementation. Furthermore, the introduction highlights the challenges faced by existing autonomous agent systems utilizing LLMs in replicating human workflows effectively and coherently. These systems often oversimplify complexities and struggle with meaningful collaborative interactions. Inspired by widely accepted Standardized Operating Procedures (SOPs) developed by humans across various domains, MetaGPT leverages SOPs to improve task decomposition and coordination among agents. The communication protocol within MetaGPT emphasizes structured communication interfaces over unconstrained natural language interfaces used in current LLM-based multi-agent frameworks. By incorporating SOPs and structured communication protocols, MetaGPT aims to enhance collaboration among various roles within a team, ultimately leading to the production of meticulously crafted software solutions. Detailed diagrams and instances showcasing the SOP workflow in MetaGPT are provided for further clarity and understanding. Overall, MetaGPT represents a promising advancement in the field of automated problem-solving through its integration of human workflows, SOPs, and structured communication protocols into LLM-based multi-agent collaborations.