Textbook of AI

MetaGPT (Hong et al., 2023) is a multi-agent system that hard-codes a software-engineering organisation. Given a one-line product brief, it produces requirements documents, design specs, code, and tests by routing through a sequence of role-specialised LLM agents.

The SOP insight

Hong et al.'s thesis is that chat is the wrong abstraction. Real organisations succeed not because employees chat freely but because they follow Standard Operating Procedures with structured artefacts:

• PM produces a Product Requirements Doc (PRD).
• Architect produces a System Design.
• Engineer produces Code.
• QA produces a Test Plan.

Each agent reads the previous artefact (not the chat history) and produces a strictly-formatted next artefact. This drastically reduces hallucination and "telephone-game" drift compared to free-form group chat.

Pipeline

User brief
  ↓
[Product Manager]   → PRD.md
  ↓
[Architect]         → SystemDesign.md, classes.json, sequence.png
  ↓
[Project Manager]   → Tasks.md (work breakdown)
  ↓
[Engineer]          → main.py, modules/*.py
  ↓
[QA Engineer]       → test_*.py
  ↓
Final code repo

Each step has a strict schema for its output (the SOP), enforced via templated prompts.

Code excerpt

from metagpt.team import Team
from metagpt.roles import ProductManager, Architect, Engineer

team = Team()
team.hire([ProductManager(), Architect(), Engineer()])
team.invest(investment=3.0)              # token budget
team.run_project("Build a 2048 game")
asyncio.run(team.run(n_round=5))

Strengths

• Better than free-form chat for software , produces complete repos with documentation.
• Reproducible, the SOP eliminates speaker-selection randomness.
• Auditable, every artefact is on disk.

Limitations

• Hard-coded SOP doesn't adapt to non-software tasks.
• Generated code typically needs human cleanup; not yet "ship to production".
• Same compounding-error problem as other long pipelines.

Modern relevance

MetaGPT is the conceptual ancestor of single-agent SWE systems like Devin, OpenHands, and OpenAI Codex (2025). Its lasting contribution is the artefact-driven pattern, agents communicate via files, not chat, which has become standard in coding agents.

Citation

Hong, S. et al. (2023). MetaGPT: Meta Programming for a Multi-Agent Collaborative Framework. ICLR 2024. arXiv:2308.00352.