Citizens Agents Experts Framework(公民·代理·专家三分法)

定义

Citizens Agents Experts Framework 是 Rachel Laycock(Thoughtworks CTO,2026-08)提出的 AI 时代价值流动框架——Citizens(任何能 turn ideas into working software 的人)+ Agents(执行:write/refactor/test/fix/iterate)+ Experts(治理:架构/安全/韧性/operability/compliance/cost)。不是 role 分类而是 value-flow 描述:AI 让 build 普及给 citizens,execution 由 agents 接管,engineering judgment 在 experts 手中杠杆放大。

三个 Buckets 的职责

Bucket做什么价值贡献
Citizens把 ideas 转化为 working software(不限于工程师)扩展创造力的可及性
Agentswrite code、refactor、generate tests、fix bugs、iterate at speed执行速度的杠杆
Experts决定 software 是否 deserve to exist in production;design guardrails、platforms、practices、feedback loopsjudgment 的杠杆

与传统角色的区别

  • 不是 role 分类:citizens / agents / experts 是value-flow 的三个阶段,不是 three job titles
  • AI 让 citizens 增加:任何人都能用 AI build——不再是 engineer 的专属
  • Agents 接管 execution:不需要 expert 亲自写每个 feature
  • Experts 不被取代:judgment 的稀缺使其从"执行 feature"翻转为"设计让 thousands features 安全的环境"

核心稀缺:Engineering Judgment

过去几十年:稀缺是 coding(engineer 难找且贵)
                  ↓
AI 时代:稀缺是 judgment
        ↓
  知道什么算好
  知道风险是否理解
  知道 works today 是否能 trust in production

Rachel 自陈:"I'm not convinced that was ever the real scarcity, but that's probably another ramble."

关键论点:"Organisations don't run on code. They run on trust."

  • 当 agents 生成大量 code 时,good design matters more, not less
  • judgment 的杠杆来自decisions 的下游影响放大(千 features 共用 platforms/guardrails)
  • experts 的工作 = 设计让 chaos 不发生的 environment

Demo 阶段 vs Production 阶段的鸿沟

Demo 阶段问题Production 阶段问题
Features workIs customer data protected?
Looks polishedWhat happens when dependency fails?
Demo impressesCan someone understand this in 2 years?
Solves stated problemWill it survive audit?
Can it cope 1000×more users?
How will we know something's wrong before customers do?

核心:以上 production 阶段问题不出现 unless experienced engineer 在场

FOSE 2026 共识(Rachel 引用)

  • FOSE 讨论"spent surprisingly little time talking about coding"
  • 主导话题:design / architecture / governance / learning / judgement
  • 典型实践:design specification → agents work overnight → review next morning
  • 共识:当 agents 可生成 lots of code quickly,good design matters more

关键数据点

  • 提出者: Rachel Laycock, Thoughtworks CTO
  • 文章日期: 2026-08-19
  • 引用事件: FOSE(Future of Software Development)
  • 类别: value-flow framework(非 role taxonomy)

与相关 concept 的关系

前提与局限性

  • 前提 1: AI 让 build 能力普及给 citizens(已是实证)
  • 前提 2: agents 可承担大部分 execution(前沿模型证据支持)
  • 前提 3: judgment 不可被 AI 完全替代(争议——Verification Tether(forward reference,未建 entity) 论证 judgment 需要 internalized mastery)
  • 边界: 三分法是 value-flow 描述而非 role boundary,实际工作中边界模糊
  • selection bias: Rachel 是 Thoughtworks CTO(enterprise consulting 视角);startup / scale-up 可能 lean agents + minimal experts
  • 未量化: experts 数量与杠杆变化未给出 metric

实施建议

  1. 重新定位 expert role:从"写每个 feature"翻转到"design guardrails/platforms"
  2. 承认 citizens 价值:让更多人 turn ideas into software
  3. 信任 agents 执行:但 expert 必须 validate production-readiness
  4. 建立 production check 清单:trust 体系而非 trust 个人

关联概念