LN: Hong et al. (2023) — MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework

Bibliographic Reference

Citation: Hong, S., Zhuge, M., Chen, J., et al. (2023). MetaGPT: Meta programming for a multi-agent collaborative framework. arXiv:2308.00352. ICLR 2024. https://arxiv.org/abs/2308.00352


Pass 1 — Bird’s Eye View (5 Cs)

CAssessment
CategorySystem proposal + empirical evaluation
ContextBuilds on LLM-as-agent (ReAct, AutoGPT), addressing coordination failures
CorrectnessEvaluated on HumanEval, MBPP, SoftwareDev benchmark. Clear ablations.
Contributions(1) Structured workflows encoding real-world SOP into multi-agent communication; (2) Role specialisation (PM, architect, engineer, QA); (3) “Code = requirement + design + code” pipeline
ClarityGood. Some implementation complexity.

Relevance: ⭐⭐⭐⭐

BMAD methodology in PUMA directly inspired by MetaGPT’s role-based approach.


Pass 2 — Content

Core Idea

MetaGPT assigns different LLM agents to structured software roles: Product Manager → Architect → Engineer → QA. Each role produces specific artifacts (PRD, design doc, code, tests). The “meta-programming” refers to encoding Standard Operating Procedures (SOPs) as agent behaviour, preventing the coordination failures of purely conversational approaches.

Key Insight for PUMA

MetaGPT’s role structure maps directly to the BMAD methodology used in PUMA: Research Analyst → Product Manager → Architect → Developer → QA → Sustainability Officer. The artifact-driven workflow (PRD → spec → implementation) mirrors SDD principles.

Limitation

MetaGPT focuses on software development tasks. Its application to ICT project management (triage, estimation, sprint planning) requires adapting the role definitions and communication protocols.


PUMA Integration

  • BMAD: MetaGPT is the theoretical foundation for the PUMA BMAD agent team → BMAD-Agent-Roster
  • Stage 5 Smart PMO: Multi-role architecture for the SwarmPM vision → Smart-PMO-Vision

MOCs