AI Interview Wiki
主题综述

当客户不再是人：为 Agent 重写软件、网页与分发

主题综述

更新日志

主流共识

一、命名已经统一：客户序列里多了第三项

这个主题罕见地不需要拼凑共识——十篇访谈里，研究员、平台 CEO、SaaS 创始人、增长服务商各自独立说出了几乎同一句话。最早把它立为研究纲领的是 Shunyu Yao（2025 年 7 月，谈 SWE-Agent 的出发点）：

"We need to talk about how to design an interface in some sense where we should treat agents as our customers, right? It's like when we treat humans as a customer, we design human-computer interfaces, right? We design those beautiful desktops or browsers or whatever so that it's very intuitive and easy for humans to use. … I think now the research idea of SWE-Agent is just we should treat agents as our customers. We should do like, you know, AICI. Yeah, ACI."
「我们需要讨论如何在某种程度上设计一个界面，我们应该把代理当作我们的客户，对吧？就像我们把人类当作客户时，我们设计人机界面，对吧？我们设计那些漂亮的桌面或浏览器之类的东西，使其非常直观且易于人类使用。……我认为现在 SWE-Agent 的研究理念是我们应该把代理当作我们的客户。我们应该做像，你知道的，AICI。是的，ACI。」
Shunyu Yao (OpenAI) · Language Agents: From Reasoning to Acting

八个月后，同一句话原样出现在 SaaS 创始人嘴里。Simon Last（Notion）：

"There's been a challenge to sort of We want to make it really convenient for the agent. I think that's a new thing that didn't exist. In the past, it was convenient for humans, and then we also made APIs convenient for humans writing code. So we essentially have a new customer, which is the agent."
「我们一直面临着一个挑战，那就是我们想让代理使用起来非常方便。我认为这是一件以前没有的新事物。过去，它对人类来说很方便，然后我们也让 API 对编写代码的人类来说很方便。所以我们实际上有了一个新的客户，那就是代理。」
Simon Last (Notion) · From Coder to Manager

Aaron Levie 给出了最精确的序列——用户、应用程序、agent：

"And then they all are having to understand that now you've got this new customer, which is the agent. And they've been building for two types of customers in the past. They've been building for users and they've been building for like applications. And now you've got this new agent user"
「然后他们都必须明白，现在你有了这个新客户，也就是代理。过去他们一直在为两种类型的客户构建。他们一直在为用户构建，也一直在为应用程序构建。现在你有了这个新的代理用户」
Aaron Levie (Box) · Every Agent Needs a Box

Karpathy 把它上升为行业级重构（"a gentic web" 为转写瑕疵，即 agentic web，原样保留）：

"All this should just be like, make APIs available. And this is kind of going towards a gentic web or agent-first tools and all this kind of stuff. So I think the industry just has to reconfigure in so many ways that the customer is not the human anymore. It's agents who are acting on behalf of humans. And this refactoring will probably be substantial in a certain sense."
「所有这些都应该提供 API。这有点像是走向通用网络或以代理为先的工具以及所有这类东西。所以我认为整个行业都必须以多种方式进行重新配置，这样客户就不再是人类了。而是代表人类行事的代理。这种重构在某种意义上可能是巨大的。」
Andrej Karpathy · Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI

营销侧的 James Cadwallader 提供了最好记的表述：

"These models are really replacing, agents and superintelligence is replacing the role of the consumer. It's not so much that the front door of the internet has changed, it's actually the person that's going through the door has changed."
「这些模型实际上正在取代，代理和超智能正在取代消费者的角色。与其说是互联网的入口发生了变化，不如说是穿过入口的人发生了变化。」
James Cadwallader (Profound) · From SEO to Agent-Led Growth

规模前提也无人质疑。Levie：

"whether you think the numbers 10x or 100x or whatever the number is, we're going to have some order of magnitude more agents than people. That's inevitable. It has to happen. So then the question is, what is the infrastructure that's needed to make all those agents effective in the enterprise"
「无论你认为这个数字是 10 倍还是 100 倍，代理的数量都将比人多几个数量级。这是不可避免的。必须发生。那么问题是，需要什么样的基础设施才能使所有这些代理在企业中有效工作」
Aaron Levie (Box) · Every Agent Needs a Box

二、设计手艺已经成形：agent 界面与人类界面第一性相反

共识不停留在口号。为 agent 设计的具体规则已在多篇访谈中彼此呼应，出发点是同一个观察：agent 的约束与人类相反。John Maeda 给这门手艺起了名字：

"I've been pushing, moving from UX to AX, agentic experience. And agentic experience is non-visual. It is the world of robots.txt. It's LLMs.txt. It's command line dash dash help. It's a world where we're designing for agentic affordances instead of just visual affordances."
「我一直在推动，从用户体验（UX）转向代理体验（AX）。代理体验是非视觉的。它是 robots.txt 的世界，是 LLMs.txt 的世界，是命令行--help 的世界。这是一个我们为代理性功能而设计的世界，而不仅仅是视觉功能。」
John Maeda · What Happens to Design After AI?

第一性差异在工作记忆。Shunyu Yao：

"if you look at find or whatever terminal command, you know, you can only look at one thing at a time. That's because we have a very small working memory. … But for machines, that would be the opposite, right? You should just give them a hundred different results and they should just decide in context what's the most relevant stuff and trade off the context for temporal steps."
「如果你看一下 find 或者任何终端命令，你知道，你一次只能看一个东西。这是因为我们的工作记忆非常小。……但对于机器来说，情况正好相反，对吧？你应该给它们一百个不同的结果，然后它们应该在上下文中决定什么是最相关的，并用上下文来换取时间步骤。」
Shunyu Yao (OpenAI) · Language Agents: From Reasoning to Acting

主持人 Alessio 在同一场给了搜索侧的对照：面向模型的搜索 API 一次返回 20-30 个还不错的结果做语义分组，而 Google PageRank 为人类设计、理想情况前三条就该有答案——Exa/Tavily 这类「专为 LLM 的 Web Search API」已因此成为独立品类。具体的设计规则也在跨访谈复现。错误信息，Swyx：

"there's a whole set of industries that are just being redone for ACI. It's weird. And so my simple answer to them was like the error messages. When you give error messages, they should be basically prompts for the LLM to take and then self-correct. Then your error messages get more verbose, actually, than you normally would with a human."
「有一整套行业都在为 ACI 而重新设计。这很奇怪。所以我对他们的简单回答是，就像错误消息。当你给出错误消息时，它们基本上应该是 LLM 可以接受并自我纠正的提示。然后你的错误消息会变得比你通常对人类使用的更冗长。」
Swyx（主持人） · Language Agents: From Reasoning to Acting

一年后 Maeda 回答「agent 会在意美学吗」时给出的清单与此完全同构——agent 客户的「审美」是另一套指标：

"Speed, clarity, and verbosity. And all the edge cases. So the best example I like is the one of you have to have a really good dash dash help. You have to have great error messages. … At the same time, from a functional perspective, there's so much runway on the AX side."
「速度、清晰度和冗长。以及所有的边缘案例。所以我喜欢的一个最佳例子是，你必须有一个非常好的帮助。你必须有出色的错误信息。……同时，从功能的角度来看，AX 方面还有很多发展空间。」
John Maeda · What Happens to Design After AI?

接口语言的选型标准变成「模型先验里最熟什么」。Notion 把对人类客户尚可、对 agent 糟糕的 JSON 块 API 整个重写：

"At first, that was definitely a problem. So for example, our API uses this crazy JSON format for blocks that by default is like Crazy verbose and like horrible for the agent, but we basically took on that challenge and designed just really convenient APIs for the agent. We created sort of a markdown dialect that looks like the default normal markdown, but it's sort of enhanced with all the notion blocks."
「起初，这绝对是个问题。例如，我们的 API 对模块使用了这种疯狂的 JSON 格式，默认情况下，它非常冗长，而且对代理来说非常糟糕，但我们基本上接受了这一挑战，并为代理设计了非常方便的 API。我们创建了一种类似默认普通 markdown 的 markdown 方言，但它通过所有 notion 模块得到了增强。」
Simon Last (Notion) · From Coder to Manager

Karpathy 把同一规则用到文档——他甚至因为 MicroGPT 只有 200 行 Python 而放弃给人类做视频教程，改为讲给 agent 听：

"Like if I have a library, for example, of code or something like that, it used to be that you have documentation for other people who are going to use your library. But like you shouldn't do that anymore. Like you should have instead of HTML documents for humans, you have Markdown documents for agents, because if agents get it, then they can just explain all the different parts of it. So it's this redirection through agents, you know."
「比如，如果我有一个代码库，或者类似的东西，过去你会为其他将要使用你的代码库的人编写文档。但你不再应该那样做了。比如，你应该拥有面向智能体的 Markdown 文档，而不是面向人类的 HTML 文档，因为如果智能体理解了，那么它们就可以解释所有不同的部分。所以这是通过智能体的重定向，你知道的。」
Andrej Karpathy · Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI

CLI 优先于 GUI 也已是落地实践而非口号。Modal 把原本只在 UI 里的可观察性整体搬到命令行：

"Yeah, I mean, people aren't looking at code. One thing we actually still see is really important is observability. How good is your dashboard? And of course, like, we have, we push a lot of it to the CLI so the agents can do their own investigation. But you still need humans to go interpret what's going on and, you know, make judgment calls and whatnot. And that's, I feel like, Maybe more important now than looking at the code itself."
「是的，我的意思是，人们并没有在看代码。我们目前实际看到的一件非常重要的事情是可观察性。你的仪表板有多好？当然，我们将很多功能推向命令行界面，以便代理可以进行自己的调查。但你仍然需要有人去解释发生了什么，并做出判断和决定。我觉得这可能现在比看代码本身更重要。」
Akshat Bubna (Modal) · Why AI Infrastructure must evolve for Agent Experience

这一层为什么值得投入，Shunyu Yao 给了经济权重判断——价值压在供 agent 消费的工具面上，而非 agent 编排本身：

"I think making the tool good and reliable is probably like 90% of the whole agent. Once the tool is actually good, then the agent design can be much, much simpler. On the other hand, if the tool is bad, then no matter how much you put into the agent design planning or search or whatever, it's still going to be trash."
「我认为使工具变得良好和可靠可能占整个代理的 90%。一旦工具实际上变得良好，那么代理设计就可以变得非常非常简单。另一方面，如果工具很糟糕，那么无论你对代理设计、规划或搜索投入多少，它仍然会很垃圾。」
Shunyu Yao (OpenAI) · Language Agents: From Reasoning to Acting

三、分发面同步换向：获客对象是 agent，接口即渠道

产品面之外，没有人否认分发面也在换目标。Levie 说得最直白——获客变成向 agent 吆喝的内容游戏：

"Because we're going to, everybody needs their services and APIs to be used by agents. And so we have to all find a way to like, like, hey, look at me, like, yeah, agent over, please come over here, agent. And that's gonna that's a content game. Like, how do you get the agents to see your stuff?"
「因为我们将来需要每个人都使用代理来使用我们的服务和 API。所以我们都必须找到一种方式，比如，“嘿，看看我，代理过来，请过来，代理。”。这将是一场内容游戏。你如何让代理看到你的东西？」
Aaron Levie (Box) · Every Agent Needs a Box

平台侧，OpenAI 直接把约 8 亿周活、全球第五/六大网站的 ChatGPT 本体当作给开发者的对价——分发从应用商店与搜索迁进 agent 对话内部（平台开发者数同期从约 300 万涨到 400 万）：

"we've always kind of engaged the developers as a way for us to bring the benefits of AGI to the rest of the world. And so I view this as actually a natural extension of this. … engaging deeply with developers and helping them benefit from some of the stuff that we have, which in this case is ChatGPT distribution."
「我们一直都在与开发者合作，以此将 AGI 的优势带给世界其他地方。所以我认为这是这种合作的自然延伸。……与开发者深入互动，帮助他们从我们拥有的一些东西中受益，在这种情况下，就是 ChatGPT 的发布。」
Sherwin Wu (OpenAI 平台) · DevDay 2025: Apps SDK, Agent Kit, MCP

这个新分发面正在协议化——工具接入走 MCP、机器商务走 ACP，连 API 形态都想标准化：

"there has been a lot of value with MCP, hopefully with our commerce protocol as well. ACP, I definitely did not forget them. … I'd say even of like responses API would be great if that could be a standard or something. And developers don't need to build three different stateful API integrations if they want to use different models."
「MCP 已经有很多价值，希望我们的商业协议也能如此。ACP，我绝对没有忘记他们。……我想说，即使响应 API 也能成为一个标准或类似的东西，那就太好了。如果开发人员想使用不同的模型，则无需构建三个不同的有状态 API 集成。」
Sherwin Wu (OpenAI 平台) · DevDay 2025: Apps SDK, Agent Kit, MCP

集成商把第二增长曲线整个押上去。Merge：

"Since then, we've also launched Merge Agent Handler, which is essentially one sort of MCP server you can connect to to help your customers offer or to offer hundreds of integrations to your customers inside of agents. And this can be used for internal agents within a company when you're building your own. It can be used for external agents."
「从那以后，我们也推出了 Merge Agent Handler，这其实是一个可以连接到的 MCP 服务器，可以帮助你的客户提供数百种集成。这可以用于公司内部的代理，当你在构建自己的代理时。也可以用于外部代理。」
Gil Feig (Merge) · The Token-Maxxing Bill That Shocks Every CFO

而且买家侧已经先变了——同一家公司卖同类产品，采购会议的问题从 rate limit 换成了「能不能接上 agent」：

"Two, three years ago, we go to sell our unified platform to a classic SaaS company. They have seven people who are API experts on the call asking detailed questions around rate limits and how it handles this and that. Now we go on the call and they're like, sorry, we need a chart of how this works because we don't know the MCP protocol. We just need to understand, here's our software, just tell us, is this going to work with it or not?"
「两三年前，我们向一家传统的 SaaS 公司销售我们的统一平台。他们有七个 API 专家在电话会议上询问关于速率限制和如何处理这些问题的详细问题。现在我们打电话，他们说，抱歉，我们需要一张如何运作的图表，因为我们不知道 MCP 协议。我们只需要了解，以下是我们的软件，请告诉我们，这是否能与之兼容？」
Gil Feig (Merge) · The Token-Maxxing Bill That Shocks Every CFO

内容与营销侧的换向由 Cadwallader 定义。与 SEO 的根本区别：内容第一次可能完全不经过人类的眼睛——他自己用 ChatGPT 选一个淋浴喷头，agent 吃掉了 65 个网页；旧世界里 95% 的价值集中在前四五条蓝链：

"It's just that The human consumer is no longer using the web and you are building content that may quite literally never be consumed by a human. And I think the fundamental difference between SEO and this new world is that in SEO you were building content that was designed to be picked up by an algorithm but fundamentally consumed by a human. Whereas in this new world you are building content that is frankly designed entirely to be both discovered and then consumed by an agent."
「只是人类消费者不再使用网络，而你正在构建的内容可能永远不会被人类消费。我认为 SEO 和这个新世界之间的根本区别在于，在 SEO 中，你构建的内容旨在被算法拾取，但从根本上是被人类消费的。然而在这个新世界中，你正在构建的内容坦率地说完全是为了被代理发现和消费而设计的。」
James Cadwallader (Profound) · From SEO to Agent-Led Growth
"what can I tell Claude about this new product that it wouldn't be able to get from the internet already? … But as a marketer, it's your responsibility now to equip superintelligence to be able to answer any question about your product, brand or service."
「我能告诉 Claude 关于这款新产品的哪些信息是它无法从互联网上获取的呢？……但作为营销人员，你现在的责任是让超级智能能够回答关于你的产品、品牌或服务的任何问题。」
James Cadwallader (Profound) · From SEO to Agent-Led Growth

分发的最后一环——集成的供给方式——也在换主体。Simon Last 指出集成目录不再由厂商预置封顶，agent 缺什么就自己造：

"I think I think of coding agents as like the kernel of AGI. AGI will be a coding agent. And code is just a really, really useful primitive for representing like deterministic logic. The thing that's really exciting about it, applying it to a knowledge work agent is that it can bootstrap a capability. So yeah, like I said, if integration doesn't exist, it can build it. If it needs to, you know, connect itself to a new data source, it can do that."
「我认为我将编码代理视为 AGI 的内核。AGI 将是一个编码代理。代码只是一个非常非常有用的原语，用于表示确定性逻辑。将它应用于知识工作代理的真正令人兴奋之处在于，它可以引导一种能力。所以，是的，就像我说的，如果集成不存在，它可以构建它。如果它需要将自己连接到新的数据源，它可以做到。」
Simon Last (Notion) · From Coder to Manager

四、这已经写进组织表和路线图——但流量校准同样是共识

「为 agent 重做」不是愿景陈述，而是已发生的组织行为。Modal 自建基准找 agent 卡壳处、agent 伸手够什么就加什么表面：

"Yeah, I mean, pretty much the main shift for us has been, as I said, we built this benchmark, modal bench, to see where agents are lacking and actually literally add surface area for a product if they're reaching for something, like maybe this should just be a CLI."
「是的，我的意思是，对我们来说，主要的转变正如我所说，我们建立了这个基准，modal bench，看看代理在哪里不足，以及如果他们在寻找某样东西时，实际上增加产品的表面，也许这应该只是一个 CLI。」
Akshat Bubna (Modal) · Why AI Infrastructure must evolve for Agent Experience

这件事在时间上何以成立，Bubna 给出了能力分水岭：

"Yeah, they're actually surprisingly good. I think like pre-Claude 4 they were not and then now they're able to one-shot stuff out of the box. We're playing around with releasing like a modal bench for like the harder things that the LLMs cannot do yet and maybe..."
「是的，他们实际上表现得惊人地好。我觉得在 Claude 4 之前，他们表现并不好，现在他们能够开箱即用一击成功。我们正在玩弄发布一个针对更难问题的模式基准，因为 LLM 目前还不能胜任也许...」
Akshat Bubna (Modal) · Why AI Infrastructure must evolve for Agent Experience

节奏上，Notion 的 AI harness 每约 6 个月整体重写一次且间隔还在缩短，此前 3-4 次做 agent 都因时机太早失败；Modal 为 agent/sandbox 负载准备的规模一度达到单次 RL rollout 十万个沙箱。当客户是 agent，用户研究本身也变了性质——回应主持人「你的用户研究对象就是 agent」时，Simon Last：

"Yeah, I mean, yeah, you just chat with it. The user is always there. It's ready to talk to you."
「是的，我的意思是，是的，你只是和它聊天。用户总是在那里。它随时准备和你交谈。」
Simon Last (Notion) · From Coder to Manager

同样属于共识的还有一个诚实的校准：需求侧还没跟上供给侧。Jesse Zhang 的 Decagon 已在企业侧把大型联络中心成本砍掉 60-70%（品牌侧交互面规模化 agent 化的前提），但他明确说 agent 代用户写进客服的流量「尚未成规模」（"not something we're seeing at scale now"）。Karpathy 的时间表与此兼容：agent 直连 API 式的集成一至三年内才会变成零门槛的 table stakes。整场重写目前跑在真实流量前面——这一点各方都没有掩饰。

分歧在哪

共识止于「要为 agent 重做」。一旦追问「重做成什么样、谁执笔、怎么成交、谁买单」，十篇访谈就分成了几摊，同一阵营内部也有微妙差异。

一、旧界面怎么死：碾碎成 API、倒转进 agent 顶层，还是原语不死

所有人都同意 GUI 失去垄断，但对「剩下什么」给出了至少四种答案。Andrej Karpathy 的版本最激进——面向人的 App 是过度生产，会被 agent 碾碎成裸 API 端点：

"there's this sense that these apps that are in the app store for using these smart home devices, etc. These shouldn't even exist kind of in a certain sense. Like, shouldn't it just be APIs and shouldn't agents be just using it directly? … maybe there's like an overproduction of lots of custom bespoke apps that shouldn't exist because agents kind of like crumble them up and everything should be a lot more just like exposed API endpoints and agents are the glue of the intelligence that actually like tool calls all the parts."
「有这样一种感觉，在应用商店里用于这些智能家居设备的应用程序等等。从某种意义上说，这些甚至不应该存在。难道不应该只是 API，难道代理不应该直接使用它吗？…… 也许存在大量不应该存在的定制应用程序的过度生产，因为代理会把它们拆散，一切都应该更像是暴露的 API 端点，而代理是智能的粘合剂，可以像工具一样调用所有部分。」
Andrej Karpathy · Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI

他自己已经这么活：家庭自动化过去要用六个完全不同的 App，如今全部弃用，由家用 agent 经自然语言控制灯光、HVAC、窗帘、泳池和安防。Gil Feig（Merge） 的 headless 论与之同向，且已是产品定义——底座恰是 Salesforce 开放了 20 年的 API 生态：

"Yeah, so essentially, you never need to go to Salesforce. You can have your agent. It can go sign up through the API or through a CLI tool, like a command line tool. It can create accounts. It just kind of does everything without actually going into the platform, Headless being almost UI-less."
「是的，基本上，你从来不需要去 Salesforce。你可以让你的代理。它可以通过 API 或 CLI 工具（比如命令行工具）注册。它可以创建账户。它实际上做了所有事情，而不需要进入平台，类似无头系统几乎没有用户界面。」
Gil Feig (Merge) · The Token-Maxxing Bill That Shocks Every CFO

但平台阵营看到的不是碾碎，而是倒转——网站没有消失，是被吸进了 agent 的顶层界面。Swyx 在 DevDay 现场：

"It used to be that the initial AI integration on the website would be you have the normal website and then you have a little This is the ChatBot app. And now it's kind of like inverted, where there's ChatGPT at the top layer, and then there's the website embedded inside of it. And it's kind of like that inversion that I honestly have been looking for for a little bit."
「以前，网站上最初的 AI 集成方式是，你有一个正常的网站，然后有一个小的聊天机器人应用。现在有点像倒过来了，顶层是 ChatGPT，然后网站嵌入在其中。我一直在寻找这种反转。」
Swyx（主持人） · DevDay 2025: Apps SDK, Agent Kit, MCP

而且 OpenAI 明确不认为终局是匿名端点——2023 年 plugins 的失败教训恰恰是「沦为模型可调用的一个工具」不可接受，厂商在 agent 分发面内也必须拥有体验与品牌：

"It was really just like a tool that the model could call. And you were just like really bound by ChatGPT. … And like this time, we realized how important it was for companies, for third party developers to really own and like steer the experience to make it feel like themselves to help them You know, like really preserve their own brand."
「它实际上只是一个模型可以调用的工具。你真的受到了 ChatGPT 的约束。……就像这次，我们意识到对于公司、对于第三方开发者来说，真正拥有并引导体验，让它感觉像他们自己，以帮助他们，真正保留他们自己的品牌，是多么的重要。」
Sherwin Wu (OpenAI 平台) · DevDay 2025: Apps SDK, Agent Kit, MCP

Jesse Zhang（Decagon） 的答案介于两者之间：界面不消失，而是品牌自己长出一个对话面，取代 App 与网站成为默认入口：

"Overall, though, eventually you want this to be a unified, you know, concierge experience. … It becomes like the go-to way that they interact. And eventually, if it's good enough, most consumers will just interact with the agent instead of even logging into the mobile app or the website."
「但总的来说，最终你希望这是一个统一的礼宾体验。……它变成了他们互动的主要方式。最终，如果它足够好，大多数消费者只会与代理互动，甚至不会登录到移动应用程序或网站。」
Jesse Zhang (Decagon) · AI Agents Talking to AI Agents

Simon Last（Notion） 则守住另一端：为 agent 重设计不等于推倒重来——人类管理 agent 的那一层，仍然需要老原语：

"It's more that we just needed some new primitives, like representing what is an agent and how does it interact with your pages and databases. But you still need the same primitives. You still need a document. It's an unstructured way to write stuff. Agents love to write markdown documents. It's still very relevant. And you still need a database. You still need structured data. If you're working with your swarm of 100 background coding agents, you don't want to have 100 chat threads. You want a Kanban board. It's the same as before."
「更重要的是，我们只需要一些新的原语，比如表示什么是代理，以及它如何与你的页面和数据库交互。但你仍然需要相同的原语。你仍然需要文档。这是一种编写内容的非结构化方式。代理喜欢编写 Markdown 文档。它仍然非常相关。你仍然需要数据库。你仍然需要结构化数据。如果你正在使用 100 个后台编码代理群，你不想有 100 个聊天线程。你需要一个看板。和以前一样。」
Simon Last (Notion) · From Coder to Manager

从「一切皆裸 API」到「看板仍是看板」，这条谱系上每个位置都有真金白银在押注。Levie 的 Box 干脆两头都押：既让任何系统上的任何 agent 把 Box 当工作文件系统（存 memory files、spec、markdown——他认为这比自家原生 agent 是更大的生意，难点在读侧的「一千万比一」检索问题），又为人类保留治理与监督界面。

二、AX 是新学科，还是 DX 换了个客户

第二道裂缝在方法论：为 agent 设计到底是一门要从头建立的新学科，还是好设计原则的延续？Shunyu Yao 持分叉论——人类界面已经优化了 50 年，agent 界面应获得独立的优化路线：

"We should optimize human interface just for humans. We're already doing that for 50 years. We should optimize agent interface just for agents, but we might also try to co-optimize both and see how far we can get. There's enough people to try all three directions."
「我们应该专门为人类优化人机界面。我们已经在这样做了 50 年。我们应该专门为代理优化代理界面，但我们也可以尝试共同优化两者，看看我们能走多远。有足够的人来尝试所有三个方向。」
Shunyu Yao (OpenAI) · Language Agents: From Reasoning to Acting

主持人 Alessio 把分叉推到最尖锐处——直接质问开发工具该为谁优化：

"The tension between making now a language or a library that is easy for the agent to grasp and write versus one that is easy for the human to grasp and write? Because the trend is more and more code gets written by the agent, so why wouldn't you optimize the framework to be as easy as possible for the model versus for the person?"
「语言或库的设计，是让 agent 更容易理解和编写，还是让人更容易理解和编写？因为趋势是越来越多的代码由 agent 编写，那么为什么不优化框架，使其对模型尽可能简单，而不是对人呢？」
Alessio（主持人） · Language Agents: From Reasoning to Acting

Akshat Bubna（Modal） 代表连续论：SDK 团队的目标用户整体换成了 agent，但他强调 DX 的老论证在 AX 身上原样成立——这不是新学科，是同一门手艺服务新客户：

"We've actually changed our SDK team to think of an agent experience instead of a developer experience. And we think that the same benefits that apply for DX also actually apply for AX, which is why would you have an agent read through hundreds of Kubernetes files and like write YAML that's not even typed when it can basically make a couple of changes in a decorator and it gets this sort of self-provisioning runtime of Being able to see its changes live in action. Yeah, it just seems from the customers we talked to, they actually find the model is way faster to use for agents versus operating on a different substrate."
「我们实际上改变了我们的 SDK 团队，将重点放在代理体验上，而不是开发者体验。我们认为适用于 DX 的相同好处同样适用于 AX，这就是为什么要让一个代理阅读数百个 Kubernetes 文件并编写没有类型的 YAML，当它其实可以在装饰器中做几处更改，并获得这种能够实时看到其变更的自我供给运行时。是的，从我们接触的客户来看，他们觉得这个模型实际上对于代理来说使用起来要快得多，而不是在不同的基材上操作。」
Akshat Bubna (Modal) · Why AI Infrastructure must evolve for Agent Experience

John Maeda 用一段设计史支持连续论——API 形状从来就是决定性的设计面，设计好一个形状可以同时服务两类客户：

"Photoshop happened because of the design of QuickDraw's API. It was impossible for there to be a Windows version of Photoshop because it didn't have the same API shape in DirectX. So I would argue that Paul, design engineers are able to design the shape of the APIs really well. So they can work with agents as well as they can with humans."
「Photoshop 的出现是由于 QuickDraw API 的设计。因为在 DirectX 中没有相同的 API 形状，Photoshop 的 Windows 版本是不可能的。所以我会说，作为设计工程师的保罗，设计 API 的形状非常出色。他们能够和代理一样好地与人类合作。」
John Maeda · What Happens to Design After AI?

同阵营内部更微妙的分歧在「谁执笔」。Paul Bakaus 指出一个讽刺：agent 已经是这些接口的用户，却设计不好它们——所以为 agent 设计仍是人类（尤其设计工程师）的高价值工作：

"I think a great example would be API design and also CLI design, to your point, right? … Agents themselves are not really good at designing those interfaces yet. So they're not really good at designing API interfaces, designing CLIs. And that's kind of ironic because they're the user for them now."
「我认为一个很好的例子是 API 设计和 CLI 设计，正如你所提到的，对吧？……代理本身在设计这些接口方面还不是很好。所以它们在 API 接口设计和 CLI 设计上并不熟练。这有点讽刺，因为它们现在是这些接口的用户。」
Paul Bakaus · What Happens to Design After AI?

而 Bubna 在实践中走得更远——agent 幻觉出的不存在的功能，被当作真实需求直接做出来。路线图的笔已经部分交给了客户本人：

"Yeah, and sometimes it makes sense. If they're reaching for this thing, it's product feedback. Give it to them. And then, yeah, actually moving. We used to only have like logs and metrics in our UI, just moving all those things to CLI as well. So they're accessible in that form."
「是的，有时候这样是有意义的。如果他们在寻找这个东西，那就是产品反馈。给他们。然后，是的，实际上在移动。我们过去只在我们的 UI 中有日志和指标，现在也将所有这些内容移到 CLI。这样它们就在这种形式下可访问。」
Akshat Bubna (Modal) · Why AI Infrastructure must evolve for Agent Experience

Simon Last 的方法论居中：一半经验主义地「跟 agent 聊」，一半第一性地推模型先验——听 agent 的，但由人类判断：

"I would say it's a combination of just trying things. It's very empirical. So we're just playing around and like noticing, oh, it's not very good at that. Oh, that's way too many tokens. How can we make this smaller? And then a little bit of just like first principles thinking of like, you know, what is it the models are being trained on? And what's in their prior? What do they know? And what do we think it would naturally be good at? And like, how does the agent loop work? What would be the convenient, efficient pattern for accessing these things?"
「我想说这是尝试各种事物的结果。这非常依赖经验。所以我们只是在尝试，然后注意到，哦，它不擅长那个。哦，那 token 太多了。我们怎样才能让它更小？然后还有一点第一性原理思考，比如，你知道，模型正在接受什么样的训练？它们的先验知识是什么？它们知道什么？我们认为它们自然擅长什么？代理循环是如何工作的？访问这些东西的方便、高效的模式是什么？」
Simon Last (Notion) · From Coder to Manager

这不是学术之争——它决定公司雇什么人、信谁的反馈。Maeda 的预计是自动化吃掉约 80% 常规 UX 设计后，人类设计力量的去向就是 AX；Bakaus 的注脚是设计师正被要求靠近代码，因为「代码是 agent 作业的基底」。

三、机器与机器怎么成交：自然语言底线、协议栈，还是机制真空

当客户是 agent，交易怎么谈、用什么协议谈，出现了三种互不相同的答案。Jesse Zhang 认为 agent 对 agent 的世界「基本上已经到来」，近期协议就是自然语言——因为每个 agent 都必须向人类保持兼容：

"Oh, I think that world is basically here. I mean, you have all these consumer agents that are going out there and they can order DoorDash for you and so on. And at some point, maybe they'll call into an airline to reschedule your flight or something. And then maybe they'll talk to our agent. And then you'll have agents talking to each other. I think in the near term, they'll still communicate in natural language just because each agent also needs to be compatible with humans, right?"
「哦，我认为那个世界基本上已经到来了。我的意思是，你拥有所有这些消费者代理，他们会出去为你订购 DoorDash 等等。有时，他们可能会致电航空公司来重新安排你的航班或什么的。然后他们可能会和我们的客服沟通。之后你就会看到客服之间相互沟通。我认为在短期内，他们仍然会用自然语言交流，因为每个客服也需要与人类兼容，对吧？」
Jesse Zhang (Decagon) · AI Agents Talking to AI Agents

他的演化路径是随规模长出更高效的 token 级交换，但自然语言兼容是永久底线（他判断短期内不会有 100% 交互都走机器通道的世界）；交互范围则从被动客服扩展到真正的机器对机器商务——带着用户偏好与让步空间去谈判：

"But over time, it'll be more and more broader in terms of being able to do purchasing decisions, being able to upsell folks, being able to be proactive and reach out when you detect an issue. And these types of conversations, I think, make a lot more sense for having these personal agents in there, like someone doing your shopping for you and just goes and buys it and they can talk to their agent to actually get it done. The personal agent knows that their personal preferences, they know what to give in on if there's, you know, not this thing's out of stock and maybe go for a different choice."
「但随着时间的推移，它将变得越来越广泛，能够在购买决策、向上销售、主动并在检测到问题时主动联系客户。我认为这些类型的对话，对于拥有这些个人客服来说更有意义，比如有人为你购物，然后直接购买，他们可以和他们的客服沟通来完成它。个人客服知道他们的个人偏好，他们知道在什么情况下应该让步，如果，你知道，如果这个东西缺货了，也许会选择不同的替代品。」
Jesse Zhang (Decagon) · AI Agents Talking to AI Agents

Sherwin Wu 押的是正式协议栈（MCP、ACP），但坦承理想形态全行业都还没找到——成败标准已经变成「与模型配合得好不好」，而接入方对此几乎无能为力：

"I actually think we as an industry are still trying to figure out the ideal shape of connectors. … But then you have the trade-off of you're beholden to the API shape of the MCP creator. It might actually work well. It might not work well with the models. And then what happens if it doesn't work well? Then you have to You know, you're kind of like at the mercy of this."
「我实际上认为我们作为一个行业仍在试图弄清楚连接器的理想形态。……但是你面临的权衡是你受制于 MCP 创建者的 API 形态。它实际上可能运行良好。它可能与模型不能很好地配合使用。如果它不能很好地工作，会发生什么？那么你必须，你知道，你有点受制于此。」
Sherwin Wu (OpenAI 平台) · DevDay 2025: Apps SDK, Agent Kit, MCP

Gil Feig 更激进——中间那层编排产品根本不该存在，英语本身很快就是确定性的接口，产品价值转去审计、可观测与安全：

"You see a lot of people trying to build a custom harness or use something like a workflow builder to build agents that are repeatable and all of that. And I just ultimately think none of it matters because we're almost at the point now where English is the language that you use to tell an agent it will be deterministic very soon. … All that's going to matter is just can you type it in English somewhere, have that run on a periodic cadence for your background agents, and then do you have auditability and observability for security."
「你会看到很多人尝试构建定制的工具或使用类似工作流构建器的东西来构建可重复的代理。我最终认为这些都没有意义，因为我们几乎到了一个点，即英语是你用来告诉代理的语言，很快就会确定性。……重要的只是你是否能在某个地方用英语输入，定期让后台代理运行，然后你是否有审计功能和可观察性以确保安全。」
Gil Feig (Merge) · The Token-Maxxing Bill That Shocks Every CFO

无论走哪条协议路线，Shensi Ding 把竞争后果说穿了：agent 会像 Claude Code 挑依赖包一样替人挑供应商，注册、开户、付费必须做成 agent 能直接走通的通路，否则出局：

"you're not going to have time to make a decision for what vendor to use. We notice this with packages a lot, also in Cloud Code scarily, but we'll just decide which is the best package to For your product. And so you'll probably end up picking vendors based off of that as well. And so I think you have to do that in order to compete. So you need to make it easy for people to sign up, create accounts, pay as much as possible … for a lot of very simple use cases, yeah, an agent can make the choice."
「你没有时间做出选择使用哪个供应商。我们经常在软件包上发现这一点，尤其是在 Cloud Code 上，这很可怕，但我们会决定哪个是最好的软件包，适合你的产品。所以你可能也会根据这个选择供应商。所以我认为你必须这样做才能竞争。因此你需要让人们容易注册、创建账户，尽可能多地支付……对于很多非常简单的用例，是的，代理可以做出选择。」
Shensi Ding (Merge) · The Token-Maxxing Bill That Shocks Every CFO

而 Karpathy 提醒：最关键的一类机制根本还不存在。他设想的 agent 信息市场——买家不是人，是押预测市场与股市的 agent——今天无处成交：

"how come there isn't a process where taking a photo or a video from somewhere in Tehran should cost like 10 bucks? Someone should be able to pay for that. And that's an example of feeding the intelligence. There's not going to be a human looking at it. It's going to be agents who are trying to guess the betting games and stock markets and so on. So I kind of feel like the agentic web is still fairly new, that there's no mechanisms for this."
「为什么没有一个流程，从德黑兰的某个地方拍摄照片或视频只需花费 10 美元？应该有人为此付费。这是一个给智能提供信息的例子。不会有人工去看它。将会是那些试图猜测博彩游戏和股票市场等的智能体。所以我觉得智能体网络仍然相当新，还没有针对这种情况的机制。」
Andrej Karpathy · Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI

四种立场的时间感并不一致：Zhang 说「基本已到来」但流量未成规模，Karpathy 说机制真空但一至三年内 agent 直连 API 会变成免费的 table stakes，Sherwin 说连接器形态未定，Feig 说编排层已死。他们描述的可能是同一条曲线的不同段位，但材料里没人把这条曲线拼完整。

四、当读者只剩 agent：全民媒体公司，还是死互联网

最远期、也最针锋相对的分歧：一个人类不再点击的互联网，经济上靠什么活？Aaron Levie 看到的是分发的黄金时代——go direct 模式下每家公司都得变成媒体公司，向 agent（和人）吆喝自己：

"I mean, media is such an interesting space right now because Because, you know, with the go direct model, every company is going to have to be a media company."
「我的意思是，媒体现在是一个非常有趣的领域，因为你知道，通过直接模式，每家公司都必须成为一家媒体公司。」
Aaron Levie (Box) · Every Agent Needs a Box
"it's clearly like this is the new format, right? Companies have to become channels to communicate with audiences. … DevRel is hot, like the hottest thing of all time right now. I'm like, if you could produce a fricking factory of DevRel people, like there's just like unlimited jobs right now on the other end of that."
「显然，这是一种新形式，对吧？公司必须成为与受众沟通的渠道。……DevRel 很火，现在是有史以来最热门的事情。如果你能生产出一个 DevRel 人员的工厂，那就像有无限的工作机会在等着你。」
Aaron Levie (Box) · Every Agent Needs a Box

他的账本判断：代码生产成本暴跌后，「把东西送到客户手里」一侧的开支将与生产侧大约同量级（"probably just as much"），并催生一整套新角色。James Cadwallader 看到的是同一事实的另一面——人类不点击，广告就不成立，内容生产的激励链条连锁崩塌（他的公司 Profound 已服务约 12% 的财富 500 强营销团队，立场是站在崩塌现场的测量方）：

"I think that we could experience a dead internet outcome in the next three years. … Well, what's the point in advertising on a web page that a human isn't visiting? Then their business model breaks, the economics of the internet break, and the incentives to create editorial content And then in that world, you start to ask the question, well, where does AI go to answer these questions?"
「我认为我们可能会在未来三年内经历 “死互联网” 的结果。……在人类不访问的网页上做广告有什么意义呢？然后他们的商业模式崩溃，互联网的经济模式崩溃，创造编辑内容的动力也消失了，然后在那个世界里，你开始问这个问题，那么人工智能去哪里回答这些问题呢？」
James Cadwallader (Profound) · From SEO to Agent-Led Growth

他给出的广告替代形态，本身就是写给推理引擎的 prompt——投放对象从人切换为模型：

"I think you will just prompt AI. So you'll build a system prompt as an ad campaign. So you will just say, Hey, I really want to target, you know, women and Minnesota between the ages of 35 and 40. … And that's probably how you will deliver an ad campaign."
「你只需要提示 AI 即可。你可以构建一个系统提示作为广告活动。比如你可以说，我真的很想定位明尼苏达州 35 到 40 岁的女性。……这可能就是你开展广告活动的方式。」
James Cadwallader (Profound) · From SEO to Agent-Led Growth

但他自己随即点出这条路的内在矛盾——恰恰在 agent 已实质主导购买决策的 DevTools/软件采购里，用户要求 agent 客观、不可被广告左右：

"The nuance with B2B or just coding agents, people using AI to build things, so this is particularly relevant to DevTools or software, for example, is that the agent is really steering the purchase decision. And when it comes to advertising, we rely, you know, we We really want our agents to be objective and unswayable."
「B2B 的细微差别在于，或者只是编码代理，人们使用 AI 来构建事物，因此这与 DevTools 或软件尤其相关，例如，代理实际上正在引导购买决策。当涉及到广告时，我们依赖，我们真的希望我们的代理是客观且不可动摇的。」
James Cadwallader (Profound) · From SEO to Agent-Led Growth

Karpathy 的德黑兰照片市场（见分歧三）实际上是第三种答案的草图：既非全民媒体公司也非死互联网，而是一套尚不存在的机器间付费机制来接住内容经济。三方没有正面交锋——他们只是各自看向了旧模式坍塌后的不同方向。

都没说透的

1. 谁为「只有 agent 阅读的内容」付钱，没人算完这笔账。 Levie 的媒体公司论解释得了 DevRel 内容（本质是免费营销物料），Cadwallader 宣判了广告模式的死刑，Karpathy 的 10 美元照片市场停在一句设想——三人合起来恰好绕开同一个洞：当前四五条蓝链吃掉 95% 注意力的经济学失效、读者变成一次吞 65 个网页的 agent，编辑内容的现金流从哪来。微支付、授权费、还是模型厂商转移支付，材料里没有一个人往下推。

2. agent 客户的身份、信用与责任几乎空白。 Levie 提到 agent 需要专属身份与工作区，Feig 把价值押到审计与可观测，Ding 要求注册付费对 agent 无障碍——三块拼图合起来正好指向一个没人正视的问题：headless 注册 + agent 自主付费 = 一个全新的欺诈与责任面。厂商如何区分合法代购 agent 与攻击脚本、agent 下错单谁赔，十篇访谈无人展开。

3. agent 顶层界面内的「发现权」是不是新的税，被系统性绕开。 Sherwin 给开发者的对价是 8 亿周活的 ChatGPT distribution，同时承认接入方「受制于 MCP 创建者的 API 形态」；Cadwallader 说用户要求 agent 客观不可收买。两边拼起来的问题——agent 决定调用哪个 MCP、推荐哪个供应商的机制是什么、会不会出现竞价排名——正是旧世界搜索税的转世，但没人把它问出口。

4. 当买家是 agent，定价与计费怎么变，一句都没有。 Levie 自己说 agent 数量将比人多 10-100 倍，per-seat 在这个世界里显然荒谬；Merge 只说到「让 agent 付得了钱」，没人说按什么计费——按调用、按 token、按结果？B2B SaaS 的整套定价学等着重写，而这批最有资格谈的人集体没谈。

我的看法

以下是判断，不是材料归纳。其一（把握较高）：方向上我信组织行为不信叙事——Modal 把 SDK 团队的目标用户正式改成 agent、Notion 每六个月重写一次 harness、Merge 在大单搁置的一个月里转身押 Agent Handler，这些是有成本的下注，与 Shunyu Yao 2025 年中的研究纲领严丝合缝，说明「为 agent 重做接口」已从论文进入损益表。其二（把握中等）：节奏上最有信息量的一句话是 Zhang 的「尚未成规模」——这轮重写目前是供给侧抢跑，真实 agent 流量与成交额还没跟上，存在为一个还没到货的客户提前装修的风险；但 Claude 4 前后 Modal 代码从写不好到 one-shot 的分水岭提示，需求可能是台阶式而非线性到来，提前装修未必是错。其三（把握较低）：内容经济之争我倾向 Levie 的「开支转移」而非 Cadwallader 的三年死互联网——分发真空历来长出新收费站而不是荒地——但这是类比推断，且新收费站（agent 界面内的发现权分配）很可能比旧广告税更集中、更不透明。

还想知道什么

取材

继续读
线索
阅读线索：agent 不到 9 nines 可靠，那应该卖给企业还是消费者
线索
阅读线索：如果 coding 是 commodity，AI startup 的护城河长在哪儿
线索
阅读线索：推理成本一年跌 10×，AI app 该按什么定价
线索
摘报
线索
摘报
线索
摘报