Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
38 tokens/sec
GPT-4o
59 tokens/sec
Gemini 2.5 Pro Pro
41 tokens/sec
o3 Pro
7 tokens/sec
GPT-4.1 Pro
50 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework (2308.00352v7)

Published 1 Aug 2023 in cs.AI and cs.MA
MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework

Abstract: Remarkable progress has been made on automated problem solving through societies of agents based on LLMs. Existing LLM-based multi-agent systems can already solve simple dialogue tasks. Solutions to more complex tasks, however, are complicated through logic inconsistencies due to cascading hallucinations caused by naively chaining LLMs. Here we introduce MetaGPT, an innovative meta-programming framework incorporating efficient human workflows into LLM-based multi-agent collaborations. MetaGPT encodes Standardized Operating Procedures (SOPs) into prompt sequences for more streamlined workflows, thus allowing agents with human-like domain expertise to verify intermediate results and reduce errors. MetaGPT utilizes an assembly line paradigm to assign diverse roles to various agents, efficiently breaking down complex tasks into subtasks involving many agents working together. On collaborative software engineering benchmarks, MetaGPT generates more coherent solutions than previous chat-based multi-agent systems. Our project can be found at https://github.com/geekan/MetaGPT

An Overview of MetaGPT: Meta Programming for a Multi-Agent Collaborative Framework

The paper "MetaGPT: Meta Programming for a Multi-Agent Collaborative Framework" introduces MetaGPT, a novel meta-programming framework designed to address limitations in current LLM-based multi-agent systems. The authors propose a method integrating Standardized Operating Procedures (SOPs) to manage complex problem-solving through a collaborative approach among specialized agents. This methodology aims to mitigate inconsistencies and errors in task execution, ultimately enhancing the performance and reliability of LLM-based multi-agent systems.

Introduction

LLM-based autonomous agents offer significant potential for enhancing human workflows. However, current systems often struggle with logic inconsistencies and inaccuracies due to cascading hallucinations when chaining LLMs. The proposed MetaGPT framework aims to overcome these limitations by adopting human methodologies, particularly SOPs, to structure multi-agent collaborations. This structured approach allows each agent to contribute domain-specific expertise, providing a more reliable and coherent solution to complex tasks.

Framework and Methodology

MetaGPT leverages an "assembly line" paradigm where tasks are broken down into specific roles assigned to various agents. Key roles within the framework include Product Manager, Architect, Project Manager, Engineer, and QA Engineer. Each role follows defined SOPs, contributing structured outputs at different stages of the task execution. This methodology ensures that intermediate results are thoroughly verified, significantly reducing errors and enhancing the overall coherence of the solution.

Agents in SOPs

The framework outlines specific role specializations, where each agent is responsible for particular aspects of a problem. For instance, the Product Manager creates Product Requirements Documents (PRDs), while the Architect designs system interfaces and flow diagrams. The Engineer generates and executes code, and the QA Engineer ensures code quality through rigorous testing. This division of labor, coupled with structured communication interfaces, enhances the efficiency and accuracy of the multi-agent collaboration.

Communication Protocol

MetaGPT employs a publish-subscribe mechanism for efficient information sharing among agents. All structured outputs are stored in a global message pool, allowing agents to access relevant information as needed. This system reduces the complexity and potential for information loss associated with one-to-one communication among numerous agents.

Iterative Programming with Executable Feedback

A crucial innovation in MetaGPT is the executable feedback mechanism. After generating initial code, the Engineer executes and debugs it, iteratively refining the code based on runtime feedback. This process significantly improves the quality of the final code, addressing issues that may be overlooked during static reviews.

Experimental Results

MetaGPT was evaluated against established benchmarks such as HumanEval and MBPP, demonstrating superior performance with state-of-the-art Pass@1 rates of 85.9% and 87.7%, respectively. Moreover, the framework was tested on a self-generated SoftwareDev dataset, encompassing diverse and complex software development tasks. MetaGPT outperformed other frameworks such as AutoGPT, LangChain, and AgentVerse in terms of executability, cost-efficiency, and human revision requirements.

Numerical Results

In the SoftwareDev evaluations, MetaGPT achieved an average executability score of 3.75 out of 4, with significantly lower human revision costs compared to alternative approaches. The framework also showed efficient token usage and execution times, underscoring its practical applicability in real-world scenarios.

Implications and Future Developments

The introduction of MetaGPT has practical implications for developing more reliable and efficient LLM-based multi-agent systems. By incorporating SOPs and structured communication interfaces, the framework addresses the challenges of maintaining coherency and reducing errors in complex problem-solving tasks. The executable feedback mechanism further enhances the robustness of the generated solutions.

Future research could explore self-improvement mechanisms, where the system learns and optimizes its SOPs and communication protocols based on past experiences. Additionally, implementing an economy-based framework for agent collaboration, similar to the Natural Language-Based Society of Mind (NLSOM), could further enhance the adaptability and efficiency of such systems.

Conclusion

MetaGPT presents a significant advancement in the meta-programming of multi-agent LLM-based systems, showcasing substantial improvements in both theoretical and practical aspects of collaborative AI frameworks. By adopting human-like workflows and robust verification mechanisms, MetaGPT sets a new benchmark for reliability and performance in automated problem solving. This work paves the way for future innovations in AI-driven multi-agent systems, emphasizing the importance of structured methodologies and iterative improvements.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (15)
  1. Sirui Hong (9 papers)
  2. Xiawu Zheng (63 papers)
  3. Yuheng Cheng (10 papers)
  4. Ceyao Zhang (11 papers)
  5. Zili Wang (52 papers)
  6. Steven Ka Shing Yau (1 paper)
  7. Zijuan Lin (2 papers)
  8. Liyang Zhou (10 papers)
  9. Chenyu Ran (1 paper)
  10. Lingfeng Xiao (1 paper)
  11. Chenglin Wu (16 papers)
  12. Mingchen Zhuge (20 papers)
  13. Jiaqi Chen (89 papers)
  14. Jinlin Wang (14 papers)
  15. Jürgen Schmidhuber (124 papers)
Youtube Logo Streamline Icon: https://streamlinehq.com