MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework
Sirui Hong, Mingchen Zhuge, Jiaqi 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, J\"urgen Schmidhuber
TL;DR
MetaGPT introduces a meta-programming framework that enhances multi-agent collaborations with human-like workflows and SOPs, significantly improving solution coherence on complex tasks.
Contribution
It presents MetaGPT, a novel framework that encodes SOPs into prompts and employs an assembly line paradigm for better multi-agent task management.
Findings
MetaGPT outperforms previous systems on software engineering benchmarks.
It reduces logical inconsistencies and hallucinations in multi-agent problem solving.
MetaGPT enables more coherent and reliable multi-agent collaboration.
Abstract
Remarkable progress has been made on automated problem solving through societies of agents based on large language models (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…
Peer Reviews
Decision·ICLR 2024 oral
1. Problem-solving: MetaGPT can solve complex tasks by using different AI agents to handle different parts of the problem. 2. Clear solutions: MetaGPT provides sensible solutions that favor tasks that require different elements to work together seamlessly. 3. Flexibility: MetaGPT can be used for different tasks, and provide multiple outputs ranging from project plans to technical documentation.
1. Methodological innovation: A key issue with MetaGPT is the lack of innovation in its methodology. While it effectively utilizes Large Language Models (LLMs) in multi-agent systems, this approach may not be significantly different from existing approaches. Or MetaGPT is different from other methods in a trivial way, I don't really see their differences as being significant and requiring exploration with greater depth. 2. Fairness of experimental comparisons: The comparison methods used to ass
See above.
See above.
1. The idea of encodes SOPs of software development into LLM-based multi-agent systems is very interesting and also pracitical to use. 2. The framework is very sound and solid, with Specialization of Roles, PRD workflow across Agents, Structured Communication for complex tasks, and a compute-efficient Message Pool mechanism with both global memory and Subscription Mechanism. It also introduces an executive feedback mechanism to enhance code generation quality during runtime. 3. MetaGPT achieves
1. Most of the experiment are on GPT4, which is expensive to access, how is the performance on the benchmarks or real development demands when used with open-source LLMs? Can you share some insight on which abilities of the LLMs matters most for the success of using this multi-agent framework and how to choose proper LLMs for use?
Code & Models
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsMulti-Agent Systems and Negotiation · Speech and dialogue systems · Natural Language Processing Techniques
MethodsFocus