基于大模型智能体的虚拟化软件开发团队

项目介绍

软件开发过程涉及到需求分析、总体设计、详细设计、编码和测试等多种活动和多种角色的开发者。随着大模型技术的快速发展，很多软件开发活动都可以通过自动化和智能化方式来实现。本项目利用大模型和多智能体技术来实现智能化软件开发全过程，通过agent模拟不同角色的软件开发人员，通过agent之间的协同完成复杂的软件开发过程。

参考文献

https://openreview.net/pdf?id=yW0AZ5wPji

ChatDev论文总结

What is ChatDev?

ChatDev, a chat-powered software-development framework integrating multiple “software agents” for active involvement in three core phases of the software lifecycle: design, coding, and testing.

Methodology

1. chat chain

ChatDev introduces chat chain to further break down each phase into smaller and manageable subtasks

  • waterfall model -> chat chain
  • chat chain -> sequential phases
  • phase -> sequential subtasks
    • coding phase -> code writing + completion
    • testing phase -> code review (static testing) + system testing (dynamic testing)
  • subtask -> extraction of solutions -> dialogue of Instructor & Assistant
  • dialogue -> loop -> instructor instructing while assistant responding
  • Agentization: inception prompting mechanism
  • Memory: short-term memory / long-term memory

2. communicative dehallucination

ChatDev devise communicative dehallucination to alleviate unexpected hallucinations, which encourages the assistant to actively seek more detailed suggestions from the instructor before delivering a formal response.

Evaluation

  • Baselines: GPT-Engineer(fundamental), MetaGPT(advanced)
  • Datasets: SRDD
  • Metrics: Completeness + Executability + Consistency = Quality

The most important: multiple agents and roles

Limitations & Solutions

  1. Firstly, the capabilities of autonomous agents in software production might be overestimated: it is crucial to clearly define detailed software requirements.
  2. Secondly, unlike traditional function level code generation, automating the evaluation of general-purpose software is highly complex: consider additional factors such as functionalities, robustness, safety, and user-friendliness.
  3. Thirdly, compared to single-agent approaches, multiple agents require more tokens and time, increasing computational demands and environmental impact: enhance agent capabilities with fewer interactions

想法

  • 针对论文提到的第二点不足，利用多模态大模型添加用户这一agent，提供对应用的界面、操作、使用等的反馈，作为增加的评判软件质量的标准并加以迭代