MetaGPT: the multi-agent framework that simulates a software company

MetaGPT, published by DeepWisdom in August 2023, encodes the Standard Operating Procedures of a software team into specialised agents: ProductManager, Architect, Engineer, QA.

Open SourceAI Open SourceMetaGPTMulti-AgentSoftware AgentsLLMAI

Origin and paper

MetaGPT was published in August 2023 by DeepWisdom, with Chenglin Wu as lead author along with a group of collaborators. The project is accompanied by the paper “MetaGPT: Meta Programming for Multi-Agent Collaborative Framework”, describing a multi-agent framework oriented to software production. The licence is MIT.

The central idea is to encode into agent interactions the Standard Operating Procedures (SOP) typical of a software company, transferring into an LLM system the roles and documentary steps of a real development process.

Roles and simulation of a software company

MetaGPT defines a set of agents with roles inspired by the organisation of a development team. The ProductManager collects requirements from the user and produces a Product Requirements Document. The Architect elaborates the technical architecture document. The ProjectManager breaks work into tasks and defines the plan. The Engineer implements software modules according to specifications. The QA runs tests and reports problems.

Each role is associated with structured prompts, specific output formats and a set of actions the agent can execute. Information transfer between roles does not happen as free conversation, but through intermediate documents — requirements, diagrams, specifications, code, test reports — that replicate the deliverables of a traditional development pipeline.

SOP as structured prompt engineering

The methodologically relevant aspect is the use of SOP as a tool to stabilise the output of multi-agent LLM systems. Typical errors of free conversations — topic drift, work duplication, inconsistencies — are mitigated by imposing a prescriptive protocol: each agent knows which inputs to expect, which outputs to produce and to whom to pass them.

Evaluation and benchmarks

The framework was evaluated with the HumanEval and MBPP benchmarks, used as reference for code generation capability, and with experiments on the end-to-end production of small applications starting from natural language requirements. The project has evolved adding features such as persistent memory between roles, external tool integration and extension to data analysis processes beyond software development.

Link: github.com/geekan/MetaGPT

Need support? Under attack? Service Status
Need support? Under attack? Service Status