Abstract: Agile development, with its iterative and rapid delivery characteristics, has became a mainstream software development model. However, it still faced challenges such as unsustainable architectural evolution, difficulty in knowledge retention, and collaboration issues stemmed from skill disparities. With the rise of large language models and agent-based technologies, an enhanced agile development framework with human-multi-agent collaboration was proposed. This framework, which took LangChain as the collaborative core and Retrieval-Augmented Generation(RAG) as the knowledge support. It achieved optimization through three core implementations: An Architecture Design Agent was introduced to generate three differentiated architecture solutions (conservative for current fit but long-term refactoring; innovative for complex implementation but future-proof; compromise balancing present and future), combined with a seven-dimensional quantitative assessment (including scalability) to assist human architects in decision-making, addressing the root causes of technical debt: modeling deficiencies and architectural evolution prediction challenges. The generation of development artifacts was automated through Agents and predefined templates. Role-based modeling of the knowledge base was conducted to clarifies agent functional boundaries and execution standards, and tasks were matched with capabilities precisely to reduce collaboration bottlenecks. Experiments based on a portal website project were demonstrated that compared with the traditional Scrum method, this framework shortened development cycles by 29.6%, reduced defect density by 46.9%, and raised user satisfaction by 16.5%. This effectively demonstrated the engineering value and potential for promotion in improving automation and knowledge-based processes, ensuring delivery efficiency, and ensuring system quality.

Key words: enhanced agile development, large language models, evolution of human-machine collaborative architecture, automated generation of artifacts, standardization of agent roles

摘要： 敏捷开发以其迭代与快速交付的特性，已成为主流的软件开发模式，但其仍面临架构不可持续演进、知识难沉淀、能力差异引发协作问题等挑战。随着大语言模型与智能体技术的兴起，本研究提出了一种以LangChain为协同中枢、检索增强生成（RAG）为知识支撑的人-多智能体协同的增强型敏捷开发框架。该框架依托三大核心实现优化：引入架构设计Agent生成三类差异化架构方案（保守型适配当前但长期需重构；创新型落地复杂但适配未来；折中型平衡当前与未来），结合可拓展性等七维量化评估辅助人类架构师决策，解决技术债务源头建模不足与架构演进预测难题；通过Agent与预设模板实现开发工件自动化生成；基于知识库角色化建模明确智能体功能边界与执行标准，精准匹配任务与能力以减少协作阻塞。基于门户网站项目的实验表明：与传统Scrum方法相比，该框架使开发周期缩短29.6%，缺陷密度降低46.9%，用户满意度提升16.5%，有效验证了其在提升自动化与知识化水平、保证交付效率与系统质量方面的工程价值与推广潜力。

关键词: 敏捷开发增强, 大语言模型, 人机协同架构演化, 工件自动化生成, Agent角色标准化