“MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework” was posted to arXiv on August 1, 2023 by a team led by Sirui Hong and Chenglin Wu, with Jurgen Schmidhuber among the authors. It models a team of language model agents on a human software company, assigning each agent a role such as product manager, architect, engineer, or QA, and having them collaborate to turn a one-line requirement into working software.
The distinctive idea is to encode Standardized Operating Procedures - the kind of structured workflows human organizations use - directly into the agents’ prompt sequences, and to organize the agents on an assembly-line paradigm where each role produces a defined artifact that the next role consumes. Passing structured documents between roles rather than free-form chat is meant to reduce the hallucination and error propagation that looser multi-agent conversations suffer. The authors reported that MetaGPT produced more coherent solutions than earlier chat-based multi-agent systems on collaborative software-engineering benchmarks.
MetaGPT is frequently cited alongside ChatDev as the 2023 work that pushed the “company of agents” metaphor furthest, and its emphasis on imposing human process structure on agent teams influenced how later frameworks thought about reliability in long, multi-step tasks.