The paper "MAGIS: LLM-Based Multi-Agent Framework for GitHub Issue Resolution" addresses the challenge of resolving GitHub issues, which involves both new code incorporation and existing code maintenance. Large Language Models (LLMs) have shown promise in code generation but struggle with repository-level issues. The authors conduct an empirical study to identify the reasons behind LLMs' limitations and propose a novel multi-agent framework called MAGIS. This framework consists of four agents: Manager, Repository Custodian, Developer, and Quality Assurance Engineer. Each agent plays a specific role in the planning and coding process, collaborating to enhance LLMs' effectiveness in resolving GitHub issues. The study reveals that the ability to locate files and lines to be modified is crucial for successful issue resolution. The authors design the framework to optimize this process, using a memory mechanism to reuse previously queried information and improve efficiency. Experiments on the SWE-bench dataset show that MAGIS outperforms popular LLMs like GPT-3.5, GPT-4, and Claude-2, achieving a 13.94% resolved ratio, an eight-fold improvement over GPT-4. The framework's effectiveness is further validated through ablation studies, which demonstrate the importance of each agent's role in the planning and coding processes. The paper concludes by highlighting the potential of LLMs in software development, particularly in resolving GitHub issues, and suggests that the proposed multi-agent framework is a promising approach to integrating LLMs into software evolution workflows.The paper "MAGIS: LLM-Based Multi-Agent Framework for GitHub Issue Resolution" addresses the challenge of resolving GitHub issues, which involves both new code incorporation and existing code maintenance. Large Language Models (LLMs) have shown promise in code generation but struggle with repository-level issues. The authors conduct an empirical study to identify the reasons behind LLMs' limitations and propose a novel multi-agent framework called MAGIS. This framework consists of four agents: Manager, Repository Custodian, Developer, and Quality Assurance Engineer. Each agent plays a specific role in the planning and coding process, collaborating to enhance LLMs' effectiveness in resolving GitHub issues. The study reveals that the ability to locate files and lines to be modified is crucial for successful issue resolution. The authors design the framework to optimize this process, using a memory mechanism to reuse previously queried information and improve efficiency. Experiments on the SWE-bench dataset show that MAGIS outperforms popular LLMs like GPT-3.5, GPT-4, and Claude-2, achieving a 13.94% resolved ratio, an eight-fold improvement over GPT-4. The framework's effectiveness is further validated through ablation studies, which demonstrate the importance of each agent's role in the planning and coding processes. The paper concludes by highlighting the potential of LLMs in software development, particularly in resolving GitHub issues, and suggests that the proposed multi-agent framework is a promising approach to integrating LLMs into software evolution workflows.