你在传统流程上接进了 ChatGPT，让工程师用 Cursor 写代码，让客服用 AI 起草回复——这会带来价值，但未必已经改了组织。先别急着给它命名，顺着一条真实工作流追问：谁决定目标？agent 得到哪些上下文？它能动到哪里？错了由谁发现、谁承担、怎么回滚？若这四件事仍按旧方式发生，比较准确的说法是 AI-enabled。把“用了 AI”直接等同于 AI Native，只会遮住真正需要重画的部分。

You have added ChatGPT to an existing workflow, let engineers write code with Cursor, and let support draft replies with AI. That can create value without yet changing the organization. Before naming it, trace one real workflow: who sets the goal, what context reaches the agent, what it is allowed to change, and who detects, owns, and rolls back an error? If those four things still work as before, AI-enabled is the more accurate name. Treating “uses AI” as AI-Native hides the part that still needs redesign.

差别往往在第一次出错时才露出来。一个 agent 给出了不合适的答复，或改坏了一段代码，流程会把谁推到台前？如果仍是原来的负责人临时补位，仍靠原来的文档拼上下文，仍要沿着原来的层级申请权限，仍在末端补救，那么组织只是跑得更快，并没有换一套运作方式。

The real difference often appears at the first mistake. An agent gives an unsuitable reply or breaks a piece of code: who does the workflow put in front of the problem? If the same owner still steps in at the last minute, assembles context from the same documents, requests permission through the same hierarchy, and repairs the damage at the end, the organization is merely moving faster. It is not operating differently.

这也把“AI 转型”放回它应有的位置。老组织当然可以从局部工作流开始重画；只是迁移成本、既有责任链和遗留系统不会因为模型更强而消失。与其给团队打成熟度分数，不如从一条高频、可回滚的流程开始试验：记录判断、上下文、权限和验证怎样变化。若结果只停在更快产出，就诚实地保留 AI-enabled；只有结构变化同时解释了新能力和新风险，这个词才有分析价值。

This also puts “AI transformation” in its proper place. An established organization can start redrawing one workflow at a time, but migration cost, inherited accountability, and legacy systems do not disappear because models improve. Rather than giving a team a maturity score, begin with one frequent, reversible workflow and record how judgment, context, permissions, and verification change. If the result stops at faster output, honestly keep AI-enabled. The term earns analytical value only when a structural change explains both a new capability and a new risk.

因此它最适合有架构权的人：创业者、能搭建独立 greenfield 单元的负责人，或愿意把局部流程当实验的人。真正的问题是：在今天的能力与责任条件下，哪一部分值得先重画，又该用什么观察证明这次重画没有把风险藏起来？

It is therefore most useful to people with architectural latitude: founders, leaders who can stand up an independent greenfield unit, or teams willing to treat a local workflow as an experiment. The real question is this: under today’s capabilities and accountability conditions, what should be redrawn first, and what observation would show that the redraw has not merely hidden its risk?

这三种视角不矛盾，是同一光谱上的三个位置——而一个真正成熟的 AI Native 组织三者都占：结构上，agent 是正式的生产单位（视角一）；运营上，每条工作流都先问 AI 能接哪一步（视角二）；到了真正想探边界的单元，还有 agent 自主运营的实验（视角三）。分清这三层，后面的讨论才不会各说各话。

These three readings are not options you pick between; they mark three positions on one spectrum, and a genuinely mature AI-Native organization occupies all three at once: structurally, agents hold formal production roles (Perspective 1); operationally, every workflow starts by asking which step AI can own (Perspective 2); and at the units actually pushing the frontier, there are experiments in autonomous agent operation (Perspective 3). Keep these three apart, or the rest of this discussion talks past itself.

本方法论的架构支柱同时回应这几个视角——"AI 优先即默认"是运营视角；"Agent 即默认工种"和"工作流即代码"在结构与运营之间；"人作为判断与责任锚"则锚定本质视角的边界：哪怕在最激进的 Agent-first 实验里，也要有个具体的人接住最终的后果。这是本方法论的押注，不是已经证成的事实——后果能不能一直锚在某个具体的人身上，恰恰是本卷最先可能被反例掀翻的一注。

The architectural pillars of this methodology respond to all three perspectives simultaneously: “AI-first as default” addresses the operational perspective; “Agent as the default worker” and “workflow as code” bridge the structural and operational; and “humans as the anchor of judgment and accountability” fixes the boundary of the ontological perspective: even in the most radical Agent-first experiment, a specific person must still catch the ultimate consequence. This is the methodology’s bet, not an established fact: whether the consequence can stay anchored to a specific person is precisely the first of this volume’s wagers a counter-example might overturn.

第一股力量，协调机器化。科斯（Coase）的交易成本（搜寻、议价、监督、执行）在 AI agent 介入后能被部分甚至整体机器化，内部协调的成本曲线因此整体下移。以前非靠层级盯着不可的活，现在靠 telemetry 和 agent guardrails 就能盯住。

The first force: coordination gets mechanized. Coase’s transaction costs (search, negotiation, monitoring, enforcement) can now be partly or wholly taken over by AI agents once they intervene, pulling the whole curve down. Work that used to need a manager watching over it now gets watched by telemetry and agent guardrails instead.

第二股力量是工作单位的反转。在传统组织里，你定义角色，工作流从角色之间的互动中涌现出来。在 AI Native 组织里，你定义工作流，角色从工作流的需求中涌现出来。这是设计逻辑的彻底反转：组织的核心文档是工作流规约，而非岗位说明书。

The second force is the work-unit inversion. In a traditional organization, you define roles and workflows emerge from the interactions between them. In an AI Native organization, you define workflows and roles emerge from the requirements of those workflows. This is a complete inversion of design logic: the organization’s core document is not a job description but a workflow specification.

还有一股力量：瓶颈从执行挪到了判断。AI 能以接近零的边际成本生成、转换、总结、执行；它做不好的事——决定什么值得做、在几个选项里怎么选、为后果担责、把组织方向稳住——变成了新的稀缺资源。组织里最值钱的人，因此从能干活的人变成了能拍板负责的人。

Another force: the bottleneck moves from execution to judgment. AI generates, transforms, summarizes, and executes at near-zero marginal cost; what it still cannot do well (deciding what’s worth doing, choosing among options, owning the consequences, holding the organization’s direction) becomes the new scarce resource. The most valuable person in the org stops being the one who can get things done and becomes the one who can call the shot and answer for it.

这三股力量合起来说的是一件事：传统组织设计一直在为错的目标优化——清晰的角色、可预测的流程、人居中协调。AI Native 设计要的是快的工作流、嵌进流程里的判断、机器居中协调。换的是优化目标本身，不是调几个参数。

Put these three forces together and one thing follows: traditional organizational design has been optimizing for the wrong target: clear roles, predictable process, human-mediated coordination. AI-Native design optimizes for fast workflows, judgment embedded in the process, machine-mediated coordination. The target itself changed; this is not a parameter tweak.

还有一个常被忽略的结构性事实加固这个判断：LLM 反转了技术扩散的历史方向。电力、计算、GPS 都是政府与企业先用、消费者后用；LLM 反过来：先触达数十亿消费者，组织反而滞后。Karpathy 2025/6 把这件事当作主题来讲[R6]，实证也跟上了：Bick-Blandin-Deming 的全国调查（NBER WP 32966 → Management Science, 2026[R7]）测得 2024 年底约 40% 的 18–64 岁人口已在使用生成式 AI——整体采纳快于 PC 与互联网同期，且由消费端驱动；而企业的正式采纳率仅 5-9%。这意味着组织重构的知识此刻在个体手里、不在制度里——AI Native 创业者不在等技术成熟，而在等组织形态追上技术。口径的诚实注脚：同一研究显示企业工作场所采纳两年达 28%，与 PC 时代速度相当——组织是相对消费浪潮慢，不是绝对慢。

There is a structural fact frequently overlooked that reinforces this judgment: LLMs reversed the historical direction of technology diffusion. Electricity, computing, and GPS all reached governments and enterprises first, consumers later; LLMs went the other way: touching billions of consumers first, while organizations lagged. Karpathy made this the theme of his 2025/6 talk[R6], and the empirical record followed: the Bick-Blandin-Deming national survey (NBER WP 32966 → Management Science, 2026[R7]) found that by end-2024, ~40% of the US population aged 18–64 were already using generative AI: adoption faster overall than the PC or the internet at the same stage, and driven by the consumer end. Formal enterprise adoption stood at only 5-9%. This means the knowledge of how to restructure organizations currently resides with individuals, not institutions. AI Native founders are waiting for organizational forms to catch up with the technology, not for the technology to mature. An honest footnote on the data: the same study shows two-year workplace adoption at 28%, comparable to the PC era: organizations are relatively slow against the consumer wave, not slow in absolute terms.

Coase 边界的当代重画The Coase Boundary, Redrawn

The Coase Boundary, Redrawn

Ronald Coase 在 1937 年提出企业之所以存在，是因为内部协调比市场协调便宜。这个回答稳定了 80 年，直到 AI Agent 出现。Williamson、Jensen-Meckling 进一步把"代理成本"加入对比，给出了"企业最优规模 = 内部追加一笔交易的边际成本 = 市场完成同一交易的边际成本"这个均衡条件。

Ronald Coase proposed in 1937 that the firm exists because internal coordination is cheaper than market coordination. That answer held for eighty years, until AI agents arrived. Williamson and Jensen-Meckling later added “agency costs” to the comparison, yielding the equilibrium condition: the optimal firm size is the point at which the marginal cost of adding one more transaction internally equals the marginal cost of completing that same transaction through the market.

AI Agent 的引入根本性地改变了这个均衡。三类成本同时下降——搜寻成本（RAG / 向量库让组织记忆秒级可达）、议价成本（Agent-to-Agent 协议如 MCP、A2A 让自动议价成为可能）、监督成本（实时观察性如 LangSmith / Helicone 让远程异步监督优于现场监督）。其逻辑结果是：传统企业的边界（哪些活动留在内部 vs 外包给市场）会大规模重画。Anysphere 以约 300 人做到 $20 亿 ARR（2026/2，人均约 $600 万，仍是 SaaS 巨头的十倍量级）、Cognition 以收购 Windsurf 前累计净烧钱不足 $2,000 万走到 $260 亿投后估值（2026/5〔TechCrunch〕）——是Coase 边界向"市场端"压缩的早期实证，不是孤立的异常值。

The introduction of AI agents fundamentally alters this equilibrium. Three categories of cost fall simultaneously: search costs (RAG / vector stores make organizational memory accessible in seconds), negotiation costs (Agent-to-Agent protocols such as MCP and A2A make automated negotiation possible), and monitoring costs (real-time observability tools such as LangSmith / Helicone make remote asynchronous supervision superior to on-site supervision). The logical consequence: the boundaries of the traditional firm (which activities stay internal, which get outsourced to the market) will be redrawn at scale. Anysphere reached ~$2B ARR with roughly 300 people (2026/2, ~$6M revenue per person, still ten times the figure for SaaS giants); Cognition reached a $26B post-money valuation (2026/5〔TechCrunch〕) on under $20M of cumulative net burn before its acquisition of Windsurf. These are early empirical evidence of the Coase boundary compressing toward the market end, not isolated outliers.

这条推演在 2025 年获得了正面的学术对话对象。NBER 工作论文《The Coasean Singularity?》（Shahidi, Rusak, Manning, Fradkin & Horton, WP 34468, 2025/11[R1]）把话说得更直接：交易成本的全部构成要素（查询价格、谈判条款、签订合约、监督履约）恰好是 AI Agent 能以极低边际成本执行的任务类型；一旦有效执行，1937 年定义的 make-or-buy 边界将显著移动。论文给存量市场画的三阶段路径是：增强人类 → 整任务替代（人类转向判断、监督与关系工作）→ 工作流围绕 Agent 能力重组，和本规约的瓶颈诊断说的是同一个道理；而它对全新市场的判断是：agent-first 市场将直接从终点状态设计——这正是「只为 greenfield 而写」这一立场的学理版本。诚实的另一半也要引：标题里的问号是作者自己打的——拥塞、价格混淆、监管构成新摩擦，"有效执行"的前提在今天尚未满足，这是理论预测，不是已实现的事实。

This reasoning found a direct academic interlocutor in 2025. The NBER working paper The Coasean Singularity? (Shahidi, Rusak, Manning, Fradkin & Horton, WP 34468, 2025/11[R1]) states it more bluntly: every component of transaction costs (querying prices, negotiating terms, signing contracts, monitoring performance) is precisely the type of task that AI agents can execute at near-zero marginal cost; if effectively executed, the make-or-buy boundary defined in 1937 will shift substantially. The paper charts a three-stage path for incumbent markets: augmenting humans → whole-task substitution (humans shift to judgment, oversight, and relational work) → workflows reorganized around agent capabilities. This path lines up with this atlas’s bottleneck diagnosis; and its verdict on greenfield markets is: agent-first markets will be designed directly from the endpoint state. That is the academic formulation of why this atlas is drawn only for greenfield. The caveats must be cited too: the question mark in the title is the authors’ own. Congestion, price confusion, and regulation constitute new frictions, and the precondition of “effective execution” has not yet been met. This is a theoretical prediction, not an accomplished fact.

但同时，一种新的成本兴起——算法封建主义（algorithmic feudalism）。当 AI 能力被 OpenAI、Anthropic、Google、Microsoft 四家巨头垄断，"AI Native 组织"实际上把核心生产要素外包给一个高度集中的供应商寡头，构成一种新的"地租依附"关系。这就是为什么"多模型架构"在架构支柱中是基础性的——它是当代 Coase 边界设计中的"主权保留"。

At the same time, a new cost is rising: algorithmic feudalism. When AI capabilities are monopolized by four giants (OpenAI, Anthropic, Google, and Microsoft), an “AI Native organization” in effect outsources its core means of production to a highly concentrated supplier oligopoly, creating a new form of rent dependency. This is why the “multi-model architecture” is foundational among the architectural pillars: it is the “sovereignty reservation” in contemporary Coase boundary design.

代理理论的扩展Agency Theory, Extended

Agency Theory, Extended

Jensen-Meckling 1976 年的代理理论建立在一个清晰的二元结构上——委托人（principal，如股东）与代理人（agent，如经理）。代理成本来自信息不对称、目标分歧、激励错位。整个公司治理结构（董事会、薪酬委员会、KPI、绩效考核）都是这个理论的工程化实现。

The Jensen-Meckling (1976) agency theory rests on a clear binary structure: principal (e.g., shareholders) and agent (e.g., managers). Agency costs arise from information asymmetry, goal divergence, and misaligned incentives. The entire apparatus of corporate governance (boards, compensation committees, KPIs, performance reviews) is the engineering implementation of this theory.

AI Agent 介入后，这个二元结构变成了三元结构——principal-agent-agent。人类经理代理股东，AI Agent 又代理人类经理。责任链多了一层，问题是这一层的责任如何分配？Air Canada 案（2024 年 BC 省 BCCRT 在 Moffatt v. Air Canada 中判决公司必须为 chatbot 承诺承担法律责任）首次明确了第一层答案——公司不能用"我们的 AI 说错了"作为免责理由。但更深的问题尚未解决——当 AI Agent 之间互相调用、互相代理（如 Anthropic Project Deal 中员工授权 Claude Opus 代为议价），责任链如何追溯？

Once AI agents intervene, this binary structure becomes a ternary structure: principal-agent-agent. Human managers act as agents for shareholders; AI agents in turn act as agents for human managers. The accountability chain gains one more layer, and the question is how accountability at that layer is allocated. The Air Canada case (2024, BC BCCRT ruling in Moffatt v. Air Canada that the company must bear legal responsibility for its chatbot’s commitments) settled the first-layer answer for the first time: a company cannot use “our AI got it wrong” as a disclaimer. But the deeper question remains unresolved: when AI agents call and proxy one another (as when an employee in Anthropic Project Deal authorizes Claude Opus to negotiate on their behalf), how is the accountability chain traced?

2025 年的组织经济学给这个三元结构补上了更激进的一块。Hadfield 与 Koh 在《An Economy of AI Agents》（arXiv:2509.01063，NBER 变革性 AI 经济学手册章节[R2]）中重新检视经典文献（Coase 的协调摩擦、Williamson 的交易成本、Grossman-Hart 的产权、Holmström-Milgrom 的代理模型）并指出：这些理论识别的企业规模上限，全部源于人类固有约束（沟通速率受限、偷懒倾向），而这些约束"似乎内在于人类、却不内在于 AI"——Agent 近即时通信，奖励函数可以被直接设计为不偷懒，于是监督与履约这两大组织开销在理论上变得不必要。论文进一步引用 Chen-Elliott-Koh 的形式模型（Journal of Economic Theory, 2023）：当 AI 压低维持异质能力的组织成本时，经济可能发生突变式相变——从大量专业化企业转向少数横跨众多行业的巨型企业。引用须保留原文的虚拟语气：这是条件性预测（"如果 Agent 确实能……"），NBER 同卷评论人 Kevin Bryan 也对变化速度提出了制度性异议——但它把"组织规模的旧均衡正在失效"从直觉升格成了可检验的理论命题。

Organizational economics in 2025 added a more radical piece to this ternary structure. Hadfield and Koh, in An Economy of AI Agents (arXiv:2509.01063, a chapter in the NBER Handbook on the Economics of Transformative AI[R2]), revisit the canonical literature (Coase’s coordination friction, Williamson’s transaction costs, Grossman-Hart property rights, Holmström-Milgrom agency models) and observe that every size limit on the firm identified by these theories derives from constraints inherent to humans (bounded communication rates, shirking tendencies), and that these constraints “seem intrinsic to humans but not to AI”: agents communicate near-instantaneously, and reward functions can be designed directly against shirking, making monitoring and enforcement (the two largest organizational overheads) theoretically unnecessary. The paper further cites the formal model of Chen-Elliott-Koh (Journal of Economic Theory, 2023): when AI lowers the organizational cost of maintaining heterogeneous capabilities, the economy may undergo a discontinuous phase transition: from many specialized firms to a small number of mega-firms spanning numerous industries. Citations must preserve the paper’s conditional mood: this is a conditional prediction (“if agents truly can…”); NBER co-volume commentator Kevin Bryan also raised institutional objections about the speed of change. But the paper elevates “the old equilibrium of firm size is failing” from intuition to a testable theoretical proposition.

这就是为什么"人作为判断与责任锚"在架构支柱中是不可妥协的——不是因为人比 AI 决策更准，而是因为只有人能承担后果。Lattice 2024 年 7 月把 AI 列为"正式员工"3 天后撤回，是因为 HR 框架要求"员工"能承担责任，而 AI 无法。这是代理理论在 AI 时代仍然成立的最深部分——责任不能委托给无法承担后果的实体。

This is why “humans as judgment and accountability anchors” is non-negotiable among the architectural pillars: not because humans make more accurate decisions than AI, but because only humans can bear consequences. When Lattice listed AI as a “formal employee” in July 2024 and reversed course three days later, the essential reason was that the HR framework requires “employees” to be capable of bearing responsibility, which AI cannot do. This is the deepest part of agency theory that still holds in the AI era: accountability cannot be delegated to an entity incapable of bearing consequences.

控制论的回归Cybernetics, Returning

Cybernetics, Returning

Stafford Beer 在 1972 年《Brain of the Firm》中提出 Viable System Model（VSM）——任何能持续运转的组织都需要五个子系统：S1 操作单元（执行）、S2 协调（避免冲突）、S3 控制（资源分配与短期优化）、S4 智能（外环境扫描与长期规划）、S5 政策（身份与价值观）。这个模型在 1980 年代被广泛尝试但最终未能主流化，因为人类无法实时执行 S2 和 S3 所需的反馈密度。

Stafford Beer proposed the Viable System Model (VSM) in his 1972 Brain of the Firm: any organization that can sustain itself requires five subsystems: S1 operational units (execution), S2 coordination (conflict avoidance), S3 control (resource allocation and short-term optimization), S4 intelligence (external environment scanning and long-term planning), S5 policy (identity and values). The model was widely attempted in the 1980s but ultimately failed to go mainstream, because humans could not sustain in real time the feedback density that S2 and S3 require.

当代 AI Agent 让 VSM 重新成为可行的组织设计语言。具体的映射是——生成式 Agent 处于 S1（执行）与 S4（探索）；telemetry 与 guardrails 处于 S2（协调）与 S3（实时控制）；人类保留 S5（身份与价值观）。这就是为什么 Anthropic 据报道的"90 天最长规划周期"能够运转——不依赖年度战略来对齐组织，而依赖 S2/S3 层的实时反馈密度（口径注：该规划周期出自高管访谈与媒体报道，未经独立验证）。Cursor、Replit、Cognition 的极速迭代节奏也是同一逻辑——VSM 在 AI 时代第一次有了可落地的实现路径。

Contemporary AI agents make VSM viable once more as an organizational design language. The concrete mapping is: generative agents occupy S1 (execution) and S4 (exploration); telemetry and guardrails occupy S2 (coordination) and S3 (real-time control); humans retain S5 (identity and values). This is why Anthropic’s reported “90-day maximum planning horizon” can function: not by relying on annual strategy to align the organization, but by relying on real-time feedback density at the S2/S3 layer (sourcing note: this planning horizon comes from executive interviews and media reports; it has not been independently verified). The hyper-velocity iteration cadence at Cursor, Replit, and Cognition follows the same logic: for the first time, VSM has an implementable execution path in the AI era.

判断稀缺性的经济学The Economics of Judgment Scarcity

The Economics of Judgment Scarcity

Daron Acemoglu 2024 年在 MIT 的研究《The Simple Macroeconomics of AI》给出了一个谨慎的测算：AI 未来 10 年累计 GDP 贡献约 1.1-1.6%（年均 ~0.05%），远低于行业普遍宣称的数倍效应。MIT NANDA 2025/7 预印报告《The GenAI Divide》测得：定制化企业 GenAI 试点在约六个月观察窗口内，95% 没有可衡量的 P&L 影响（报告自述口径：52 家访谈 + 153 份问卷 + 300+ 公开部署，非同行评议；媒体流传的"150 访谈 + 350 问卷"系二手转述，详见 R14）。这两组材料不能证明一个普遍瓶颈，却提出了一个值得检查的问题：当执行被加速时，价值是否卡在流程其余部分，还是成本、证据与时间窗口尚不足以看见效果？

Daron Acemoglu’s 2024 MIT research The Simple Macroeconomics of AI offers a cautious estimate: AI’s cumulative GDP contribution over the next ten years will be approximately 1.1-1.6% (annual average ~0.05%), far below the multiples commonly claimed by the industry. The MIT NANDA 2025/7 preprint report The GenAI Divide measured that 95% of customized enterprise GenAI pilots showed no measurable P&L impact within an approximately six-month observation window (report’s self-stated method: 52 interviews + 153 surveys + 300+ public deployments, non-peer-reviewed; the widely circulated “150 interviews + 350 surveys” is a second-hand restatement; see R14). These materials do not prove a universal bottleneck. They pose a question worth checking: when execution accelerates, is value stuck elsewhere in the workflow, or are cost, evidence, and the time window still too weak to show an effect?

"判断稀缺"在经济学里有正主文献——出自 Agrawal、Gans 与 Goldfarb 的 prediction-vs-judgment 框架，并非 Acemoglu。他们 2018 年的形式模型（NBER WP 24626；同行评审版刊于 Information Economics and Policy, 2019[R3]）给出三个本规约直接继承的结论：① AI 降低的是"预测"这一特定任务的成本：预测是决策的输入，不是决策本身；② 判断被形式化定义为"目标函数无法被描述或编码时人类所行使的能力"：并非所有人类判断都与 AI 互补，更便宜的预测以相反方向影响不同类型判断的回报；③ 委托定理：即便人类参与能产出更优决策，人类仍会理性地把部分决策完全委托给机器：Agent 在严格优于人类之前就获得完全自治，是模型内最优选择的推论，并非失误。

口径校正：AGG 这条"目标函数无法被编码时的人类能力"是按 AI 能力反向定义的移动靶——AI 每强一次，"无法编码"的外延就收缩一寸。所以本方法论不把判断锚在"AI 做不到"上，而锚在一个正面、不随能力移动的定义：判断＝为不可逆后果承担责任的授权节点（与全站责任锚同义；再往下退一层是"谁定义什么算好的后果"，见下方断裂点）。它押的是"收缩有下界"——这是可错的赌注，不是定理。还要说清这注押在哪一层：判断变稀缺不是前提，是执行变便宜之后被推出来的结论；而把执行推向便宜的同一股力量，没理由在判断门口自动停步。所以本卷最先该盯的改判信号，不是外面谁来反驳，而是那条一向只压执行的成本曲线，开始稳定地压低拿主意本身的成本——真到那天，要改写的就不止哪根支柱，而是这卷立论的前提本身。

2025 年他们把"判断"进一步拆开（《The Economics of Bicycles for the Mind》, NBER WP 34034[R4]）：机会判断（识别什么值得启动）在模型中恒为认知工具的互补品——AI 提升而非侵蚀它的价值；收益判断（知道在给定状态下采取何种行动）只在工具不过度削减人类努力时才互补；而实现技能被建模为认知工具的替代品。一句话翻译：AI 吃掉实现、抬高机会判断、对收益判断态度暧昧——这恰好是"人即判断锚点"与"操作者即编排者"两个世界观的经济学坐标。同一谱系里，Gans 的《AI as Strategist》（NBER WP 33650, 2025[R5]）从控制权角度独立推出一个和 FIG 5.1 判断锚点地图讲同一个道理的结论：授予战略家正式控制权的增量价值随其可信度单调递减——所以组织应当逐域（domain-by-domain）而非统一地分配 AI 的控制与影响力：判断密集域人类主导、数据丰富域 AI 主要靠透明推理产生影响力而非权威。三篇全是理论模型而非实证：引用它们，是给"判断稀缺"找可对话、可证伪的学术对象，不是宣称已被证明。

“Judgment scarcity” has a canonical economics literature: the prediction-vs-judgment framework of Agrawal, Gans, and Goldfarb (AGG), not Acemoglu. Their 2018 formal model (NBER WP 24626; peer-reviewed version published in Information Economics and Policy, 2019[R3]) yields three conclusions this atlas inherits directly: ① AI reduces the cost of “prediction” as a specific task: prediction is an input to decisions, not the decision itself; ② judgment is formally defined as “the capability humans exercise when the objective function cannot be described or encoded”: not all human judgment is complementary to AI; cheaper prediction affects the returns to different types of judgment in opposite directions; ③ the delegation theorem: even when human involvement yields better decisions, humans will rationally delegate some decisions entirely to machines. Agents acquiring full autonomy before they strictly outperform humans is a corollary of optimal choice within the model, not a failure.

A calibration: AGG’s “the capability humans exercise when the objective function cannot be encoded” is a moving target defined in reverse from AI’s reach: each time AI strengthens, the extension of “cannot be encoded” contracts an inch. So this methodology does not anchor judgment on “what AI can’t do” but on a positive definition that does not move with capability: judgment = the authorized node that bears responsibility for irreversible consequences (synonymous with the site-wide accountability anchor; one layer deeper it retreats to “who defines what counts as a good consequence,” see the fracture below). Its bet is that “the contraction has a lower bound,” a fallible wager, not a theorem. It is worth naming which layer the wager sits on: judgment’s scarcity is not a premise but a conclusion pushed out once execution gets cheap, and the same force that drove execution cheap has no reason to halt politely at judgment’s door. So the first revision signal this volume should watch is not who argues against it from outside, but the moment that cost curve, which only ever pressed on execution, starts steadily pressing down the cost of making the call itself. The day it does, what gets rewritten is not one pillar but this volume’s own founding premise.

In 2025 they decomposed “judgment” further (The Economics of Bicycles for the Mind, NBER WP 34034[R4]): opportunity judgment (identifying what is worth starting) is modeled as a persistent complement to cognitive tools: AI raises rather than erodes its value; payoff judgment (knowing which action to take given a state) is complementary only when the tool does not excessively reduce human effort; implementation skill, by contrast, is modeled as a substitute for cognitive tools. In one sentence: AI consumes implementation, elevates opportunity judgment, and is ambivalent about payoff judgment. That is precisely the economic coordinate of “humans as judgment anchors” and “operators as orchestrators.” In the same lineage, Gans’s AI as Strategist (NBER WP 33650, 2025[R5]) independently derives from a control-rights angle a conclusion that lines up with the FIG 5.1 judgment-anchor map: the incremental value of granting a strategist formal control rights decreases monotonically with their credibility, so organizations should allocate AI control and influence domain-by-domain rather than uniformly: in judgment-dense domains humans lead; in data-rich domains AI exerts influence primarily through transparent reasoning rather than authority. All three are theoretical models, not empirical studies. Citing them gives “judgment scarcity” a scholarly interlocutor that can be engaged and falsified; it is not a claim that the thesis has been proven.

顺着这条线读，Acemoglu 的谨慎测算与 AGG 的微观模型提出同一个条件性诊断：当 AI 压低某类执行成本，而目标、验证与责任仍未同速自动化时，组织会更频繁地遇到判断问题——决定什么值得做、怎样在备选中选择、谁来承担后果、怎样维持方向（第 3 节把它写成工作模型）。这不等于 AI 永远不能参与判断，也不等于责任永远只能落在自然人身上；它描述的是当前法律、社会与治理安排下，授权与追责尚未被无损外包的现实。KPI 因而不该只量执行产出，还应量决策质量、验证覆盖和风险代价。Anysphere 人均创收约 $600 万（按 2026 年 ~300 人与 $2B ARR 同期口径计）是一条有趣但不足以解释机制的观察；人均 ARR 一旦成为目标，同样会被博弈。Goodhart 定律不给本方法论豁免权。

Reading along this line, Acemoglu’s cautious estimate and AGG’s microeconomic model suggest a conditional diagnosis: when AI lowers the cost of a class of execution while goals, verification, and responsibility are not automated at the same rate, organizations encounter judgment questions more often: what is worth doing, how to choose among alternatives, who bears consequences, and how to maintain direction (Section 3 writes this as a working model). This does not mean AI can never participate in judgment, nor that responsibility can forever rest only with natural persons. It describes a current legal, social, and governance arrangement in which authorization and accountability have not been losslessly outsourced. KPIs should therefore measure more than execution output: decision quality, verification coverage, and risk cost matter too. Anysphere’s approximate $6M revenue per head (using ~300 people and $2B ARR for 2026 on a same-period basis) is an interesting observation, not sufficient evidence for a mechanism; once revenue per employee becomes the target, it can be gamed too. Goodhart’s Law grants no exemption to this methodology.

这几股力量各自瓦解了一个旧假设。它们合起来指向什么？把散落的线索装配成一个命题、检验它能否承重，是下一节的工作。

Each of these forces dismantles one old assumption. What they point to together is the work of the next section: assembling the scattered threads into a proposition and testing whether it can bear weight.

这次推导要做的，是把争论放到可检查的地方，不是宣布一条终局结论。Coase 1937 年问：市场明明有效，公司为什么还要存在？答案是交易成本——市场协调有摩擦，于是把一部分协调收进组织内部自己做（完整论述见 第 2 节）。Simon 1947 年问：人既是有限理性的，组织又如何可能？答案是结构——用层级和流程补上单个大脑不够用的带宽。AI 没有取消这两问；它改变了其中一部分成本。于是真正要检验的是：哪些成本真的下降了，下降后新的排队点出现在哪里，谁仍要为结果签字，不只是“AI 会不会替代组织”这个问题。

This derivation does not announce an end-state. It moves the argument to things we can inspect. Coase asked in 1937: if markets are efficient, why do firms exist at all? The answer is transaction costs: market coordination has friction, so some coordination is brought inside the organization (full treatment in Section 2). Simon asked in 1947: if humans are boundedly rational, how is organization even possible? The answer is structure: hierarchy and process patch the bandwidth a single mind runs short of. AI does not cancel either question; it changes some of their costs. The test is therefore which costs actually fell, where new queues appear, and who still signs for the outcome — not simply “will AI replace organizations?”

先把这副模型当成一次有条件的押注。它押三件事：判断仍与人的责任相连；足够多的上下文可以外化并可靠流动；生成能力的增长快过验证与承担后果的能力。任一条件反转，模型都该降级或改写。特别是“判断带宽”只指人的带宽，不应偷偷延伸成对所有智能的永恒断言。

Treat this model first as a conditional bet. It bets on three things: judgment remains tied to human accountability; enough context can be externalized and moved reliably; generation improves faster than verification and consequence-bearing. Reverse any condition and the model should be weakened or redrawn. In particular, “judgment bandwidth” names human bandwidth, not an eternal claim about every form of intelligence.

第 2 节 给出了测量结果：绝大多数企业 GenAI 试点对真实盈亏没有可衡量的影响。本章回答"为什么"：因为这些组织把 AI 部署在节点上（让某个人、某个环节更快），而组织的吞吐量是图的属性——由依赖链的拓扑、协调成本的曲线、决策队列的深度决定。工具改变节点的速度，改不了图的形状。

Section 2 surfaced the measurement: the great majority of enterprise GenAI pilots show no measurable profit-and-loss impact. This chapter answers “why”: these organizations deploy AI on the nodes (making a person or a step faster), yet an organization’s throughput is a property of the graph, set by the topology of its dependency chains, the curve of its coordination cost, and the depth of its decision queues. Tools change the speed of a node; they do not change the shape of the graph.

黄益贺（2026）描述过这条定律的一个具象版本（从业者观察，非受控研究）：风投机构给每个分析师配上最强的 AI，让十个 agent 并行生成十份研究报告，然后所有报告仍由同一个人按顺序阅读、由同一个投委会按周期审议。生产端并行了，消费端依旧串行，两到四周的流程几乎没有缩短。他对此有一句精确的总结："AI 不是自动让组织变快，它只是把串行瓶颈照得更亮。"——可并行的部分被 AI 加速之后，真正拖慢组织的环节会暴露得前所未有地明显：慢的不是写报告，是谁来读、谁来判断、谁来拍板。个体层面的证据同样刺眼：METR 2025 年的随机对照试验中（16 名资深开源维护者、246 个任务、各自深耕多年的百万行级代码库），开发者使用 AI 工具后实际慢了 19%，却自认为快了约 20%。研究者明确警告此结果不应外推到陌生代码库或从零构建的场景——但它至少确认了一件事：叠加层面的收益可能远比体感小，而体感本身不可信。

Huang Yihe (2026) described a concrete instance of this law (a practitioner observation, not a controlled study): a VC firm equips every analyst with the most powerful AI and lets ten agents generate ten research reports in parallel. All the reports are then still read sequentially by the same person, and reviewed on the same committee cadence. The production side parallelized; the consumption side remained serial. A two-to-four-week process barely shortened. His precise summary: “AI doesn’t automatically make an organization faster - it just illuminates serial bottlenecks more brightly.” Once the parallelizable parts are accelerated by AI, the stages that truly slow the organization become exposed as never before: the bottleneck is not writing reports, but who reads them, who judges them, who decides. Individual-level evidence is equally stark: in METR’s 2025 randomized controlled trial (16 experienced open-source maintainers, 246 tasks, each working in a million-line codebase they had cultivated for years), developers using AI tools were actually 19% slower, yet estimated themselves to be about 20% faster. The researchers explicitly cautioned against extrapolating to unfamiliar codebases or greenfield builds, but the trial nails one thing down: overlay-level gains may be far smaller than perceived, and perception itself cannot be trusted.

判别一个问题属于哪一层：工具层（单点执行慢，AI 直接可解）、流程层（顺序可重排，流程再造可解）、结构层（瓶颈由组织的存在方式本身产生，只能重构）。以下这些瓶颈全部位于结构层。每一个都按同一格式解剖：机制（传统组织为什么必然产生它）、为什么加 AI 无效（叠加悖论的具体形态）、AI Native 重构（映射到心智模型 M.01-M.05 与架构支柱）、检验信号（你的组织是否已经越过它）。

Diagnosing which layer a problem belongs to: tooling layer (a single node executes slowly; AI can fix this directly); process layer (sequence can be reordered; process reengineering can fix this); structural layer (bottleneck is generated by the organization’s very mode of existence; only restructuring can fix this). The following these bottlenecks all reside at the structural layer. Each is dissected in the same format: mechanism (why the traditional organization inevitably produces it), why overlay fails (the specific form of the overlay paradox), AI Native restructure (mapping to mental models M.01-M.05 and the architectural pillars), test signal (whether your organization has already cleared it).

这些案例合起来不是一张分类表，而是一组彼此制衡的证据。Anthropic 与 Anysphere 证明"从零架构"走得通，Inflection 与 Adept 提醒你它不保证存活；Shopify 证明大组织能转，Klarna 的回摆划出替代的边界；Air Canada 判例把责任钉在公司身上；Midjourney 则防止把"小团队大杠杆"错记在 AI 名下。AI 时代变化太快，任何分类体系都会比它引用的数据先过时——所以这里不建分类学，每家公司只说明一件事，完整口径都登记在附录。

Taken together, these cases are not a classification table but a set of mutually checking pieces of evidence. Anthropic and Anysphere prove that architecting from zero works, while Inflection and Adept remind you it guarantees no survival; Shopify proves a large organization can turn, while Klarna’s swing-back marks the boundary of replacement; the Air Canada ruling pins accountability on the company; and Midjourney keeps “small team, huge leverage” from being misfiled as an AI invention. Things now move too fast for any taxonomy to outlive the data it cites, so instead of a classification scheme, each company here shows one thing, with full sourcing registered in the appendix.

范围提醒 · 过渡态样本。这九家全部诞生于人类组织向 AI-Native 迁移的过渡期。它们佐证的是"人类组织在 AI 冲击下如何变形"，与"生而 AI-Native 的组织如何运转"是两个不同命题——后者要等够老的原生样本，眼下无人能提供，终态验证因此是一笔仍然未决的公开赌注。

Scope note · transition-state samples. All nine were born during the migration of human organizations toward AI-Native. They attest to “how human organizations deform under AI,” a different proposition from “how a born-AI-Native organization runs”: the latter awaits samples old enough to exist, which no one can yet supply, so terminal-state validation remains an open, unsettled bet.

这些数字对照同光谱上 Cursor 量级的 $2B ARR 并不大——但结构信号极强：它们证明组织的下限已经脱离人数约束，正如光谱中段的 Anysphere 与 Anthropic 证明了人均产出的上限同样脱离了直觉约束。两端是同一命题（T1）在参数空间两侧的不同解。

Against the $2B ARR of a Cursor-scale company on the same spectrum, these figures are not large, but the structural signal is very strong: they prove the lower bound of organization has already detached from a headcount constraint, just as Anysphere and Anthropic in the middle of the spectrum proved that the upper bound of output per person has likewise detached from intuitive limits. The two ends are different solutions of the same proposition (T1) on opposite sides of the parameter space.

但左端有诚实的注脚，必须连同样本一起读。其一，孤立判断没有冗余——一人公司在放大判断密度的同时放大了判断者的盲区，没有第二个节点做交叉校验，决策质量的衰减是结构性的。其二，单一供应商即生存风险——单一模型供应商一纸条款变更就能掐断命脉，多模型架构（第 7 节 支柱 04）在 N=1 时是命脉，而非支柱。其三，样本有偏——左端样本几乎全部来自低边际成本的数字产品，外推到重资产、强监管或物理交付，目前没有证据。一人公司是 T1 的极限解与试金石，而非普遍处方。它在光谱上的全部意义，是把"组织必须是很多人"这个隐含假设，永久地变成了一个待论证的命题。

But the left end carries honest footnotes that must be read together with the samples. First, isolated judgment has no redundancy: while a one-person company amplifies judgment density, it amplifies the judge’s blind spots too, and with no second node for cross-verification the decay of decision quality is structural. Second, a single vendor is a survival risk: a single model vendor can sever the lifeline with one clause change, so a multi-model architecture (Section 7, pillar 04) is, at N=1, not a pillar but the lifeline. Third, the sample is biased: the left-end samples come almost entirely from low-marginal-cost digital products, and there is currently no evidence for extrapolating to capital-heavy, heavily regulated, or physically delivered businesses. The one-person company is T1’s limiting solution and litmus test, not a universal prescription: its whole meaning on the spectrum is to turn the buried assumption that “an organization must be many people” permanently into a proposition awaiting proof.

AI 时代的创业不只是"用 AI"，更是"用 AI 创业"。Anthropic 2026 年发布的《The Founder's Playbook: Building an AI-Native Startup》系统化了这条新路径——AI 重新定义了传统创业生命周期的每一阶段。Idea 阶段的核心从抢先建造，变成了抵御"过早建造"的诱惑；MVP 阶段不再只是写代码，而是积累持久上下文；Launch 阶段的重心从抢市场，转向"消化技术债 + 释放创始人"；Scale 阶段也告别堆人，转而把领域专长编码为不可复制的护城河。

Building in the AI era is “founding a company with AI,” not just “using AI.” The Founder’s Playbook: Building an AI-Native Startup, published by Anthropic in 2026, systematizes this new path: AI redefines every stage of the traditional startup lifecycle. The Idea stage shifts from building first to resisting the temptation to build too early. The MVP stage is no longer just writing code; it is accumulating compounding context. The Launch stage moves from a land grab to the start of “paying down technical debt and freeing the founder.” The Scale stage leaves headcount growth behind, encoding domain expertise into a moat no one can copy.

把这四个阶段叠在一起看，会发现一个核心规律——创始人的位置在每一阶段都向"系统设计者"上移。这条角色演化曲线本身，比任何工具都更接近 AI Native 方法论的本质。

Layer the four stages on top of one another and one rule appears: the founder’s position rises toward “system designer” at every stage. That curve of role evolution is itself closer to the essence of the AI Native methodology than any tool.

把四个阶段叠在一起看，AI Native 创业的核心节奏不是"加速建造"——这是浅层的误读。核心节奏是"加速验证 + 持续积累上下文 + 系统化释放创始人 + 把专长编码为护城河"。每一阶段，创始人的位置都在向"系统设计者"上移：Idea 阶段从"建造者"上移到"验证设计者"；MVP 阶段从"代码作者"上移到"上下文工程师"；Launch 阶段从"决策者"上移到"系统设计者"；Scale 阶段从"内部执行者"上移到"对外角色"。

Layer the four stages together and the core rhythm of AI Native founding is not “build faster”; that is the shallow misreading. The core rhythm is “validate faster, keep accumulating context, systematically free the founder, and encode expertise into a moat”. At every stage the founder’s position rises toward “system designer”: in the Idea stage from “builder” to “validation designer”; in MVP from “code author” to “context engineer”; in Launch from “decision maker” to “system designer”; in Scale from “internal executor” to “external-facing role”.

这条角色演化曲线，就是 AI Native 方法论的终极产物。它解释了为什么 Anysphere、Cognition、Replit 这样的公司能用数十到数百人创造十亿到数百亿美元估值——他们不只是"用 AI 的传统创业团队"，他们的创始人在每一阶段都把更多执行交给系统、把更多判断留给自己。媒体报道中 Anthropic 的"Hive Mind"工作方式（90 天最长规划、Slack 长文替代会议、Project Vend 让 Claude 独立运营——前两条出自报道与访谈，未经独立验证；第三条是官方公开的负结果实验[R18]）是同一逻辑的极端版本。AI Native 方法论的真正胜利不在工具——而在创始人本身的角色演化。如果你走完这四个阶段，发现自己还在做 Stage 1 时做的工作，那么方法论失败了，不论 ARR 多高。

That curve of role evolution is the ultimate product of the AI Native methodology. It explains why companies like Anysphere, Cognition, and Replit can create valuations of one to tens of billions of dollars with tens to hundreds of people. They are not just “traditional startup teams that use AI”; at every stage their founders hand more execution to the system and keep more judgment for themselves. Anthropic’s “Hive Mind” way of working as reported in the press (planning horizons of up to 90 days, long Slack write-ups in place of meetings, and Project Vend letting Claude run a business on its own; the first two come from reporting and interviews and are not independently verified, while the third is an officially published negative-result experiment [R18]) is an extreme version of the same logic. The real win of the AI Native methodology is in the evolution of the founder’s own role, not in the tools. If you finish these four stages and find yourself still doing the work you did in Stage 1, the methodology has failed, however high the ARR.

这个路径刻意做得最低限度——它不是关于"如何成为下一个 Anysphere"，是关于"如何不在前 6 个月陷入 AI Theater"。多数 AI Native 转型在第 3 个月就崩盘了，原因不是技术问题，是第一性原理没清就开始堆工具栈。建立了原则、代码化了流程、有了上下文层和可观测性——剩下的是时间和复利的工作。没建立这些底层，再多的 Agent 也只是 Theater。

This path is deliberately kept minimal: it is not about “how to become the next Anysphere” but about “how not to fall into AI Theater in the first six months.” Most AI Native transformations collapse by month 3, not for a technical reason but because they start stacking tool stacks before the first principles are clear. Once you have established the principles, turned process into code, and built the context layer and observability, what remains is the work of time and compounding. Without those foundations, any number of agents is still just Theater.