Recently, remarkable progress has been made in automated task-solving throughthe use of multi-agents driven by large language models (LLMs). However,existing works primarily focuses on simple tasks lacking exploration andinvestigation in complicated tasks mainly due to the hallucination problem.This kind of hallucination gets amplified infinitely as multiple intelligentagents interact with each other, resulting in failures when tacklingcomplicated problems.Therefore, we introduce MetaGPT, an innovative frameworkthat infuses effective human workflows as a meta programming approach intoLLM-driven multi-agent collaboration. In particular, MetaGPT first encodesStandardized Operating Procedures (SOPs) into prompts, fostering structuredcoordination. And then, it further mandates modular outputs, bestowing agentswith domain expertise paralleling human professionals to validate outputs andreduce compounded errors. In this way, MetaGPT leverages the assembly line workmodel to assign diverse roles to various agents, thus establishing a frameworkthat can effectively and cohesively deconstruct complex multi-agentcollaborative problems. Our experiments conducted on collaborative softwareengineering tasks illustrate MetaGPT's capability in producing comprehensivesolutions with higher coherence relative to existing conversational andchat-based multi-agent systems. This underscores the potential of incorporatinghuman domain knowledge into multi-agents, thus opening up novel avenues forgrappling with intricate real-world challenges. The GitHub repository of thisproject is made publicly available on: https://github.com/geekan/MetaGPT
URL
Affiliations
Abstract
Translation (by gpt-3.5-turbo)
Summary (by gpt-3.5-turbo)