[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-co11ap5e-multi-agent":3},{"id":4,"slug":5,"name":6,"summary":7,"stack":8,"background":12,"responsibility":13,"workflow":14,"architecture":21,"highlights":22,"retrospective":27,"repositoryUrl":28,"liveUrl":29,"featured":30},7,"co11ap5e-multi-agent","co11ap5e_multi_agent（多 Agent 软件研发团队）","类 MetaGPT 的多 Agent 软件研发系统，PM \u002F 架构师 \u002F 程序员 \u002F 测试 \u002F PMO 五个角色协作完成需求 → PRD → 设计 → 编码 → 测试全流程，支持自动修 bug 与设计回退。",[9,10,11],"Python","多智能体","角色编排","用多个专职角色 Agent 模拟软件研发团队：PM 写 PRD、架构师做设计、程序员编码、测试工程师验证，PMO 纯代码编排并控制回退，测试失败自动回退修 bug，多轮失败回退改设计。","独立完成系统设计与实现。\n\n- 设计 5 角色分工：PM \u002F Architect \u002F Coder \u002F Tester \u002F PMO\n- 实现测试失败自动回退 Coder 修 bug（最多 3 轮），超限回退 Architect 改设计\n- 用 WorkspaceContext 共享文件与状态，Agent 间不直接互调\n- 文件操作限制在 workspace 内，命令执行带超时保护\n- 按 assistant+tool 消息组整体截断记忆，避免半轮消息",[15,16,17,18,19,20],"接收用户需求","PM 分析需求并输出 PRD","Architect 输出技术设计","Coder 逐文件编码并自验证","Tester 写测试并运行","失败则回退 Coder \u002F Architect，通过则交付","基于自研 react_agent 框架，不依赖 LangChain \u002F AutoGen；PMO 纯代码编排；WorkspaceContext 共享上下文；工具沙箱限制文件操作范围。",[23,24,25,26],"5 角色协作覆盖研发全流程","自动回退：3 轮修 bug，超限改设计","共享上下文不直接互调，职责清晰","工具沙箱与超时保护","多智能体系统的难点不是让多个 Agent 说话，而是定义清晰的职责边界、共享状态和回退策略。PMO 用代码控制流程，比让 LLM 自由对话更可靠。","https:\u002F\u002Fgithub.com\u002FS1rryNut\u002Fco11ap5e_multi_agent",null,true]