MetaGPT: Multi‑Agent Collaboration and Agent Capability Enhancement

Introduction – The article presents MetaGPT, an advanced multi‑agent framework released by DeepWisdom in June 2023, which quickly gained attention, achieving 40K GitHub stars and top rankings on GitHub Trending. It has been recognized with an oral presentation at ICLR 2024.

MetaGPT Development and Impact – MetaGPT integrates human‑centred standard operating procedures (SOPs) into agent development, enabling roles such as product manager, architect, project manager, developer, and tester to collaborate in a waterfall‑style software development process. This design translates high‑level user intents into structured, multi‑file code, improving code generation success rates (e.g., HumanEval Pass@1 from 67% to 85.9%).

Multi‑Agent Collaboration and Self‑Improvement – The framework supports diverse scenarios, including software creation (e.g., a 2048 game with ~20 lines of code), virtual town simulations, and a Werewolf game where agents exhibit memory management, reflection, and experience‑driven learning. Memory modules evaluate importance, recency, and relevance, allowing agents to retrieve and combine contextual information dynamically.

Practical Applications and Advantages – MetaGPT outperforms comparable frameworks in project‑level code generation tasks, delivering higher completion rates and more stable outputs. Quantitative experiments on HumanEval, MBPP, and various Kaggle datasets demonstrate significant performance gains, with the Data Interpreter achieving up to 0.95 average score across benchmarks.

Exploring Agent Capability Enhancement – The authors discuss challenges such as hallucination in code generation and propose adding code execution feedback, verification, and white‑box testing to improve reliability. They also introduce a Data Interpreter that combines code execution, tool integration, and dynamic planning to handle evolving data‑science tasks.

Tool Integration and Use Cases – Tools are treated as reusable code snippets (functions, classes, scripts) that agents can recommend, rank, and compose on‑the‑fly. This approach enables agents to embed local code, perform multimodal tasks (e.g., stable diffusion), and automate web‑scraping or email drafting, creating a closed‑loop tool evolution process.

Q&A Session – The article concludes with a Q&A covering differences between Data Interpreter and CodeInterpreter, advantages of MetaGPT over AutoGen, and strategies for tool selection and evaluation, emphasizing ensemble, debate, and review mechanisms.

Overall, the presentation showcases MetaGPT’s contributions to multi‑agent collaboration, agent memory and reasoning enhancements, and practical AI applications across software engineering and data science.