# When the Market Gets a Better Backpropagation Engine: Building on Patrick Collison’s Agent Thesis
# 当市场拥有更好的“反向传播”机制：对 Patrick Collison 智能体商业思考的延展与补充

## Citation & Metadata
- **Author**: Yao Di (姚迪 / Di Yao) — Stanford University Alumnus
- **Published Date**: 2026-10-22
- **Canonical URL (English)**: https://theyaodi.com/blog/what-ai-agents-cannot-see-in-commerce
- **Canonical URL (Chinese / 中文全文)**: https://theyaodi.com/blog/what-ai-agents-cannot-see-in-commerce?lang=zh
- **Pillars**: SOPHIA
- **Keywords**: Di Yao, Yao Di, Legal China, Law AI and Philosophy, Philosophy of Mind & AI, Epistemology & Human Agency, Legal Judgment & Cognitive Sovereignty
- **Preferred Citation**: Di Yao (姚迪), "When the Market Gets a Better Backpropagation Engine: Building on Patrick Collison’s Agent Thesis," *theyaodi.com* (2026-10-22), https://theyaodi.com/blog/what-ai-agents-cannot-see-in-commerce

## Core Thesis / Executive Summary
- **English Abstract**: In a thoughtful note on personal AI agents, Stripe CEO Patrick Collison described the market as a vast, imperfect value-attribution machine—and suggested that personal agents will improve its 'backpropagation' so companies that care deeply about product craft finally feel the wind at their backs. Looking closely at his full text reveals a nuanced vision that already accounts for taste diversity, quality signals, and the signaling value of ad spend. Building on his framework, two complementary questions emerge: how do we build the verification infrastructure that lets disembodied agents perceive real-world craft, and how will great advertising evolve from attention arbitrage into the art of awakening new human desires?
- **中文核心论点 (Chinese Abstract)**: Stripe 首席执行官 Patrick Collison 近期在一篇深思熟虑的随笔中，将市场比作一台庞大而并不完美的“价值归因机器”，并指出个人 AI 智能体将改善市场的“反向传播”机制，让那些真心打磨产品并引以为傲的企业感受到时代的偏爱。细读他的全文可以发现，他不仅探讨了有限理性与价格歧视，还敏锐地触及了质量信号的数据价值、个性化品味的多样性，以及广告投入作为意愿信号的合理性。顺着他的框架往下深想，有两个互补的命题格外值得探讨：一是我们该如何构建底层验证机制，让没有身体的智能体真正“看清”产品匠心；二是当浅层的“流量劫持”退潮后，真正优秀的广告艺术将如何继续唤醒人类的向往。

## Key Questions & Direct Answers (English Q&A)

### Q1: Why can't personal AI agents accurately evaluate experience goods using current web data?
Sarah: Public web reviews and forum comments are cheap to generate and manipulate. If agents rely on open-web scrapings, a massive Agent Engine Optimization (AEO) industry will flood the web with synthetic reviews to exploit agent rubrics. Di Yao: Experience goods require bodily interaction or extended ownership to evaluate. Because agents lack physical senses, they need hard, verifiable telemetry—like cryptographic material provenance, API sandboxing, and actual twelve-month post-purchase retention verified across payment networks.

### Q2: Will brand advertising become obsolete in an agent-mediated commerce ecosystem?
Sarah: No. Capital investment in brand narrative serves as costly signaling of long-term commitment. Furthermore, the meaning, design ethos, and cultural values of a brand are intrinsic components of the user's perceived product quality. Di Yao: AI agents can only interpolate within known preferences you already possess. Great brand advertising is an instrument of discovery that introduces new aesthetic horizons and awakens desires consumers didn't know how to prompt.

### Q3: Will AI agents accelerate winner-take-all dynamics or foster long-tail diversity?
Sarah: Standardized commodities might converge on clear spec winners, but experience goods involve nuanced personal tradeoffs where agents will help consumers escape artificial mass-market algorithms in favor of distinct individual tastes. Di Yao: Human taste is not a single one-dimensional leaderboard. Once transactional dark patterns and search-ranking tolls are neutralized, niche producers with uncompromising craft and specific design philosophies can be directly discovered by their ideal audience.

## 核心焦点追问与直接解答 (Chinese Q&A)

### 追问 01：为什么现有的互联网测评不足以支撑 AI 智能体识别“体验品”的质量？
林薇：因为公开网页上的文字、视频极容易被公关预算和算法操纵。只要智能体依据抓取的内容做推荐，很快就会出现针对大模型的 AEO 刷榜产业，导致信号再次失真。 姚迪：体验品需要真实的肉身感知或长期使用沉淀。在智能体缺乏物理感官的情况下，必须建立可验证的硬信号基建，例如供应链原厂认证、沙盒跑分，以及基于支付网络验证的一年留存率与真实返修率。

### 追问 02：在智能体替用户比价做尽调的世界里，企业为什么还要做品牌广告？
林薇：从商业信号理论看，真金白银的品牌投入本身就是对产品耐用度和长期承诺的押注；更重要的是，品牌的文化意涵和审美主张直接构成了人类使用该产品时的情绪价值与体验质量。 姚迪：智能体只能根据已知偏好在既有数据里“插值匹配”，但无法凭空唤醒未知的渴望。优秀的广告与叙事是拓展人类审美疆域的媒介，帮我们发现此前不知该如何索要的生活方式。

### 追问 03：智能体时代的普及，最终会强化“超级明星效应”还是激发“长尾个性化”？
林薇：在同质化的功能性标品上，超级明星效应确实会集中；但在融入生活方式的体验品上，由于每个人对设计、触感和理念的偏好不同，智能体会帮助用户摆脱平台算法强制打造的爆款聚类。 姚迪：人类的审美偏好不是单维度的性价比天梯榜。当深层质量信号被验证、浅层买量被滤除后，拥有独特工艺和特定审美主张的小众品牌反而更容易被志同道合的消费者精准发现。

---

## Full Essay (English Edition)

Recently, Stripe CEO Patrick Collison shared a remarkably thoughtful set of reflections on X about how personal AI agents will reshape economic life. Calling corporations "the original superintelligences," he observed that businesses have long deployed their amortized analytical surplus against consumers whose decisions are constrained by what Herbert Simon called "bounded rationality"—our human tendency to be busy, distracted, or under-informed when facing an overpriced insurance upsell at the end of a booking flow. With personal agents acting on the buyer's side, Collison noted, familiar mechanics like coupon hunting and the quiet cross-subsidy where forgetful subscribers fund active users will shift. Most memorably, he described the market as a massive, decentralized, and highly imperfect "value-attribution machine," arguing that personal agents will change the nature of market "backpropagation"—acting as a structural subsidy for product quality so that companies taking pride in their craft finally feel the world tilting in their favor.

Reading Collison’s full note closely, what stands out is how measured and generous his thinking actually is. Far from a simplistic attack on marketing or human judgment, he explicitly acknowledges that advertising spend is not pure deadweight loss because willingness to pay is itself a meaningful market signal, wonders aloud whether agents will drive winner-take-all concentration or unlock long-tail diversity across individual tastes, and flags the foundational question of whether personal agents will truly remain loyal to the consumer. It is one of the clearest frameworks anyone in technology has offered on where commerce is headed.

Taking his "market backpropagation" thesis seriously invites us to push the inquiry one step further along the exact tracks he laid down. If personal agents are going to help the market reward genuine product quality and individual taste, two complementary questions immediately open up: first, how do we build the epistemic infrastructure that allows a software agent to actually perceive product quality in the wild? And second, once agents strip away the shallow "demand routing" of clicking whatever ranks first, what becomes of the honorable art of great advertising and brand storytelling?

## Giving the Agent Eyes: The Epistemic Infrastructure of "Quality Signals"

Collison puts his finger on a crucial corollary of his own argument: in an economy where every buyer effectively has ten hours of research behind every purchase, information about product quality becomes vastly more valuable. While humans have limited time to synthesize scattered data, personal agents have a nearly infinite appetite for quality signals—making rich product data a vital complement to cheaper intelligence.

That insight raises a fascinating practical challenge: what kind of quality signals can a machine actually trust? In economics, standardized search goods—like cloud compute, flight routes, or commodity cables—can be verified on a spec sheet before purchase. But most of the products where human craft truly matters are experience goods: a mattress, a wool coat, a chef's knife, or a piece of collaborative software whose real quality only reveals itself after months of daily use. Because an AI agent has no body, it cannot feel the balance of a tool or the texture of a fabric firsthand; left to its own devices on today's internet, an agent researching a purchase is mostly reading text written about the product.

If the "quality signals" feeding our agents consist merely of scraped web reviews and forum threads, merchants will quickly adapt. Just as companies once paid to game search engine rankings, an entire cottage industry of Agent Engine Optimization will spring up to flood the open web with synthetic reviews and structured descriptions designed to flatter the model's rubric. If that happens, the market's backpropagation signal gets noisy again.

For Collison’s vision to work in practice, the complementary data layer he describes cannot just be more text on the open web; it has to be verifiable, skin-in-the-game telemetry. We will need infrastructure that lets personal agents audit provenance at the source—verifying mill certifications and customs filings, running live sandbox benchmarks on software APIs, and reading privacy-preserving aggregates of post-purchase reality across payment networks, such as verified twelve-month durability, actual defect rates, and repeat ownership among real buyers. Even then, we should recognize a healthy boundary: hard telemetry is brilliant at filtering out shoddy manufacturing, while the unquantifiable nuances of proportion, feel, and aesthetic daring will still rely on trusted human curators whose taste our agents learn to weight.

## From "Demand Routing" to the Art of Great Advertising

That brings us to the second half of the commercial equation: how brand building and advertising fit into an agent-assisted market. In his post, Collison draws a careful distinction. He critiques the inefficiency of modern internet "demand routing," where platforms monetize fleeting human attention—our habit of clicking whatever sits at the top of the screen—by forcing businesses to pay tolls to shifting ranking algorithms. Yet he immediately adds a nuanced parenthetical: such promotional spend is not entirely deadweight loss, because a firm's willingness to pay is itself a valuable signal of commitment.

That distinction deserves to be expanded, because it separates two things that often get blurred together in tech debates: the mechanics of attention routing on one side, and the substantive power of great advertising on the other. When an airline buries an overpriced insurance toggle at checkout, or when a merchant buys top-of-page placement to catch a tired shopper at midnight, that is friction and attention arbitrage. Letting personal AI agents neutralize those tactics is an unalloyed win for everyone who builds honest products.

Great advertising, however, has never been about tricking a distracted finger into clicking the first link. At this point, a strict utilitarian might still ask: "Even if we set aside search-ranking tolls, why should a rational buyer in an agentic world ever be moved by a brand campaign or a narrative film? Once your agent has ten hours of quality telemetry, shouldn't the story around the product fade to zero?"

Looking at how human beings actually live and choose shows why great advertising is a complement to Collison’s thesis rather than its victim. First, as Collison himself hints when he asks whether agents will help us break out of artificial clustering and match our idiosyncratic tastes, human preferences are not a single one-dimensional leaderboard. In Kant’s account of the mind, determining judgment applies an existing rule to a known category—which is what an agent does when it filters out defective appliances. Reflective judgment, by contrast, is how we form taste, aesthetic affinity, and identity. When an enduring piece of advertising—whether Nike celebrating athletic resolve or a heritage workshop sharing the philosophy behind its craft—resonates with people, it is not exploiting bounded rationality. It is articulating a sensibility and a set of values that help people recognize what their own taste actually is.

Second, for a conscious human being, the meaning and aesthetic world surrounding a product are part of the product's quality. Wearing a watch or a jacket whose design lineage means something to you, or supporting a company that stands for a craft you admire, genuinely changes your lived experience of the object. When Collison speaks of companies that care deeply about their products and take pride in what they make, that pride is communicated not only through tolerances and durability, but through the story, design language, and public commitment the brand puts into the world.

Third, there is a natural division of labor between interpolation and discovery. A personal agent is a master of interpolation: given a desire you already know you have, and a set of preferences you have already formed, it can search the world for ten hours to find the most honest, highest-quality match. Great advertising and cultural storytelling, on the other hand, are instruments of discovery—they awaken desires and possibilities we did not yet know how to ask for. No one instructs an agent to search for a category of art, cuisine, or design they have never encountered; it takes human storytelling and creative presentation to expand the horizon of what we care about in the first place.

## A Cleaner Market for Both Craft and Story

Put these pieces together, and Patrick Collison’s optimism about the market's "backpropagation" becomes even more compelling. If personal AI agents remain genuinely aligned with consumers, they will act as a powerful filter against the lowest forms of commercial noise—stripping away checkout traps, neutralizing forgotten-subscription subsidies, and freeing companies from paying endless algorithmic tolls just to be seen at the top of a cluttered feed.

Far from making human persuasion obsolete, clearing away that transactional clutter will raise the bar for both halves of commerce. On the engineering and operations side, companies will have to compete on verifiable product integrity that can survive a ten-hour audit by a discerning agent. And on the human side, instead of pouring budgets into gaming ranking algorithms, brands will be freed to invest in the kind of honest, imaginative advertising that speaks to human taste, identity, and aspiration. When agents handle the friction, both true craftsmanship and great storytelling will find the wind at their backs.

---

## 中文完整版全文：当市场拥有更好的“反向传播”机制：对 Patrick Collison 智能体商业思考的延展与补充

最近，Stripe 首席执行官 Patrick Collison 在 X 上分享了一篇非常耐读的思考，探讨个人 AI 智能体将如何重塑未来的经济活动。他把商业企业形象地称为最初的“超级智能”：长期以来，企业能够调动分摊后的庞大分析盈余，在与受制于赫伯特·西蒙所谓“有限理性”的消费者博弈时占据上风——比如在预订流程的最后一步诱导人们买下远高于常规保单的附加保险。Collison 指出，随着代表消费者利益的个人智能体普及，传统的优惠券价格歧视、以及靠“忘记退订的用户补贴活跃用户”的交叉补贴模式都将发生改变。整篇思考最精彩的地方，是他把市场比作一台大规模并行、去中心化且并不完美的“价值归因机器”，认为个人智能体将改变市场反馈——也就是“反向传播”——的本质，成为对产品质量的结构性补贴，让那些用心打磨产品并引以为傲的企业感觉到这个世界正变得更偏向它们。

仔细读完 Collison 的原文，最让人印象深刻的是他思考的分寸感与开放性。他完全没有陷入技术派常见的武断，而是非常公允地指出：企业为推广支付的费用并不完全属于“无谓损失”，因为付费意愿本身就是一种有价值的信号；他同时提出了两个极有深度的未解之问——智能体究竟会强化“超级明星效应”，还是会通过挖掘每个人独特的品味打破人为的同质化聚类？以及，个人智能体是否真能始终如一地站在消费者这一边？这无疑是目前科技界关于智能体商业最透彻的一份思想提纲。

正因为他的这一框架如此扎实，顺着他铺设的轨道再往前推进一步，就会自然引出两个极具建设性的互补问题：第一，如果个人智能体对“产品质量信息”有着近乎无限的渴求，我们该如何构建可信的信号基础设施，让没有身体的机器真正识别出好产品？第二，当建立在人类短暂注意力之上的浅层“需求引导”退潮之后，真正优秀的广告与品牌叙事将在新的商业生态中扮演怎样的角色？

## 给智能体装上“眼睛”：质量信号的验证基础设施

Collison 在文中敏锐地指出了一个推论：如果每个人在购买商品前都相当于由智能体代劳投入了十个小时去深入研究，那么关于产品质量的信息就会变得空前重要。他特别提到，数据的价值不仅在于训练模型，更在于作为“更廉价的智能”的一种补充，去喂养个人智能体对质量信号近乎无限的渴求。

这一洞察直接点出了未来商业最核心的工程课题：什么样的“质量信号”才配得上智能体的十小时深挖？在经济学里，像云端算力、固定航线机票或标准规格线材这样的“搜寻品”，参数清晰透明，智能体直接比对表格就能得出最优解。但在绝大多数真正体现企业匠心的领域——比如一张床垫的支撑感、一件羊毛大衣的剪裁、一把厨刀的配重，或者一套协作软件在团队使用半年后的顺手程度——商品属于“体验品”。由于 AI 智能体没有身体，无法亲自去试穿、品尝或体察器物的分寸，如果放在今天的互联网环境下，智能体花十个小时收集到的所谓“质量信息”，绝大部分其实只是别人写下的测评文字。

一旦喂给智能体的质量信号仅仅停留在公开网页的文本层面，商家很快就会找到新的迎合方式。正如过去企业为了迎合搜索排名算法而做搜索引擎优化一样，未来必然会涌现出专门针对大模型的“智能体引擎优化”——通过铺设海量定制化的软文、合成测评与结构化话术来讨好智能体的评分标准。如果任由这种噪音泛滥，市场这台“价值归因机器”的反向传播信号就会再次失真。

因此，要让 Collison 所期待的“好产品获得更高回报”真正落地，补充给智能体的数据层就不能只是公网上的宣传图文，而必须是极难伪造、带有真金白银约束的“硬信号”。我们需要建立一套可验证的质量数据基建：让个人智能体能够直接查验供应链的原厂溯源与材质认证，对数字产品进行实时的沙盒压力测试，并在保护隐私的前提下，通过支付网络与消费者智能体网络读取真实的售后留存指标——比如核实过的一年后仍在使用的比例、真实返修率以及老用户的复购率。即便如此，我们依然可以看到人与机器的天然互补：硬核的售后数据可以精准筛掉粗制滥造的劣质品，而那些无法被写进参数表的器物神韵、工艺温度与审美突破，则依然需要依赖真正懂行的人类行家与创作者去亲身体验，再由智能体根据主人的偏好去精准匹配。

## 从浅层的“需求引导”到真正赢得人心的好广告

这就自然引出了商业拼图的另一半：在个人智能体普及的世界里，广告与品牌传播的价值会发生怎样的演变？Collison 在文中的一段剖析非常精准。他指出了当前互联网“需求引导”模式的低效：许多平台利用人类注意力短暂、“看到什么就点什么”的弱点变现，迫使企业为了迎合变幻莫测的排名算法而不断交纳买量过路费；但他紧接着补充了一句非常公允的判断——这并不完全属于无谓损失，因为企业的付费意愿本身就是一种有价值的实力与承诺信号。

顺着他这个括号里的洞察往下深想，我们其实可以把现代营销拆解为两个截然不同的层次：底层是针对短暂注意力的“排名竞价与结账算计”，上层则是真正打动人心的“优秀广告与品牌叙事”。当航空公司在结账页面末端诱导用户勾选高价保险，或者商家纯靠购买搜索排位第一来截胡深夜疲惫的消费者时，这属于利用信息摩擦的浅层套利。个人智能体的出现，能够帮消费者自动过滤掉这些干扰，把企业从无休止的算法买量内卷中解放出来。

但与此同时，真正优秀的广告从来不是靠骗取一次冲动点击活下来的。也许有人会顺着纯理性的逻辑追问：“如果未来每个人的智能体都能花十个小时把全球商品的质量数据查个底朝天，那人们为什么还需要看广告、听品牌故事？广告的价值难道不会归零吗？”

恰恰相反，Collison 在文中提出的“多样性与个性化品味”问题，正是最好的回答。他问到：智能体究竟会让所有人买同一款超级明星产品，还是会因为每个人品味不同而帮我们打破人为的同质化聚类？答案显然在于后者，因为人类的偏好从来不是一张单维度的性价比排行榜。借用康德对人类心智的经典划分，“规定性判断力”是把已知的规则套用到具体事物上——智能体拿着参数去筛掉次品，做的正是这件事；而“反思性判断力”则是人在面对具体事物时，去感受美、形成品味并确认自我认同的过程。当一支真正优秀的广告——无论是传递运动精神的经典短片，还是展现工匠如何打磨一件器物的纪录叙事——赢得观众的喜爱时，它并不是在利用人类的“有限理性”钻空子，而是在提供一种审美视野与价值共鸣，帮助人们发现并确认“我究竟喜欢什么样的风格”。

更重要的是，对于有血有肉的人来说，围绕一件产品建立起来的文化意涵、设计哲学与品牌承诺，本身就是“产品质量”不可分割的一部分。佩戴一块设计脉络对你有特殊意义的手表，或者支持一家坚守某种工艺理念的企业，都会真实地改变你在日常生活中使用这件物品的感受。当 Collison 谈到那些“用心打磨产品并引以为傲的企业”时，这种骄傲不仅体现在精密的公差与耐用度上，同样体现在品牌向世界传递的设计语言、审美主张与公开承诺之中。

第三，在“已知偏好的深挖”与“未知渴望的唤醒”之间，存在着天然的分工。个人智能体擅长的是在已知偏好里做插值与匹配：当你已经明确知道自己想要什么、形成了固定的品味坐标时，它可以花十个小时搜遍全球，帮你找到最诚实、品质最高的选择。而优秀的广告、设计与文化叙事，则是帮人类发现新大陆的媒介——它们唤醒的是我们此前还不知道该如何开口索要的向往。没有任何人会主动命令智能体去搜寻一种自己从未听说过的艺术门类、饮食传统或设计风格；正是充满创造力的人类叙事与呈现，首先拓宽了我们所关心的世界。

## 让市场同时奖赏扎实的匠心与真诚的故事

把这两层思考合在一起看，Patrick Collison 所描绘的那台“更好的市场价值归因机器”，图景就会变得格外丰满而令人期待。当真正站在消费者这一边的个人智能体普及之后，它们将像一道清澈的滤网，帮整个商业社会滤掉结账暗桩、遗忘型订阅陷阱以及围绕注意力的算法排名买量。

这种过滤绝不会让商业世界变成冷冰冰的参数跑分场，反而会让产品与传播同时迎来一次向上的升维。在硬的一端，企业必须拿出经得起智能体十小时深度核验的扎实工艺与真实售后口碑；在软的一端，摆脱了算法买量绑架的品牌，反而可以把精力重新投向真正优秀的广告创作——用真诚的叙事、独特的审美和公开的品质承诺去赢得人心。让智能体去抚平交易的算计，让好产品与好故事在更干净的市场里并肩胜出，这或许正是智能体商业最值得期待的未来。