• Code Large Language Models (CodeLLMs) have proficiency in code generation but struggle with complex software tasks.
  • Recent works like ChatDev and MetaGPT introduce multi-agent frameworks for software development.
  • AgileCoder, proposed by FPT Software AI Center, mimics real-world software development using Agile Methodology.
  • AgileCoder emphasizes dynamic adaptability, structured development in sprints, and collaboration among agents.
  • Key innovation: Dynamic Code Graph Generator creates a Code Dependency Graph to model code relationships.
  • AgileCoder outperforms existing methods in benchmarks like HumanEval, MBPP, and ProjectDev.

AgileCoderは、Agile Methodologyからインスピレーションを得た新しいフレームワークであり、複雑なソフトウェア開発タスクにおいてCodeLLMsよりも優れた性能を示しています。主な革新点は、Code関係をモデル化するCode Dependency Graphを作成するDynamic Code Graph Generatorです。AgileCoderは、HumanEval、MBPP、ProjectDevなどのベンチマークで既存の手法を凌駕しています。

元記事: https://www.marktechpost.com/2024/08/10/researchers-at-fpt-software-ai-center-introduce-agilecoder-a-multi-agent-system-for-generating-complex-software-surpassing-metagpt-and-chatdev/