多 Agent 质量治理系统Multi-agent quality-governance system· 2026
TeachOps
多 Agent 教学设计预审与质量治理系统Multi-agent lesson-design review and quality governance
让取证、修订、稽核围绕同一条证据链协作,缺证据时主动阻断下游。Evidence, revision, and audit collaborate around one traceable trail, with downstream work blocked when evidence is missing.
案例摘要Case summary
目标用户Audience
需要审查教学设计质量、核对课程依据并保留审批责任的导师和教师培训团队。Mentors and teacher-training teams that need to review lesson quality, verify curriculum grounding, and retain approval responsibility.
我的角色Role
负责场景收敛、Agent 职责、证据阻断规则、验收标准和正常/异常流程验证;代码与协作配置由 AI 编程工具完成。Owned the scenario scope, agent responsibilities, evidence-gating rules, acceptance criteria, and normal/exception flow validation; code and collaboration configuration were produced with AI coding tools.
交付结果Outcome
完成正常流程与缺证据异常流程的 live 验证,产出证据包、修订稿、审计报告和真人审批记录。Completed live validation of both the normal and missing-evidence flows, producing evidence packets, revisions, audit reports, and a human approval record.
产品洞察Product insight
多 Agent 的价值不在于让多个模型同时输出,而在于明确职责、共享结构化证据,并在关键前提缺失时阻止错误继续传播。The value of multi-agent systems is not making several models produce text at once. It is separating responsibilities, sharing structured evidence, and stopping bad assumptions from propagating.
项目简介Overview
TeachOps 面向师范院校和教师培训导师,把教学设计预审拆成 Evidence、Design、Audit 多 Agent 协作流程。每条建议都绑定证据条目,关键输入缺失时不生成看似完整的结果。TeachOps is a pre-review system for lesson designs used by teacher-training teams. It coordinates Evidence, Design, and Audit agents around structured evidence, and refuses to generate a polished-looking result when key inputs are missing.
技术栈Tech stack
结果证据Evidence
README 记录四 Agent 的职责边界,以及 Evidence、Design、Audit 三个可单独测试的 Skill。The README documents four agent boundaries and three independently testable Skills for evidence, design, and audit.
正常流程已真实产出证据包、修订稿、审计报告和导师附条件批准。The normal flow produced a live evidence packet, revision, audit report, and conditional mentor approval.
缺少关键课标输入时,Evidence Agent 返回 BLOCKED,后续 Design 与 Audit 不被调用。When a required curriculum input is missing, the Evidence Agent returns BLOCKED and downstream Design and Audit agents are not called.
关键决策Key decisions
如果证据不足仍让下游 Agent 继续生成,最终结果会看似完整但无法追溯。If downstream agents continue after evidence is missing, the final output can look complete while remaining untraceable.
把证据包 READY 状态作为下游调用前置条件,缺证据直接 BLOCKED。Make a READY evidence packet a precondition for downstream calls; missing evidence returns BLOCKED.
多 Agent 容易互相越权,导致审计者替代导师做最终判断。Multi-agent workflows can blur authority and let an auditor replace the mentor’s final judgment.
为 Manager、Evidence、Design、Audit 分别设定输入、输出和不可承担的职责边界,最终审批保留给真人。Define inputs, outputs, and forbidden responsibilities for Manager, Evidence, Design, and Audit, while keeping final approval human.
想看完整的代码和实现细节?Want the full code and implementation details?
在 GitHub 上查看View on GitHub