Vercel CEO Guillermo Rauch：拆分模型与 Agent 的战斗

凭借其让开发者无需管理服务器即可部署智能体的云基础设施，Vercel 已悄然成为 AI 软件领域最核心的公司之一。该公司目前每天处理 600 万次部署，其中一半由编码 Agent 触发，每天有超过 1 万亿 Token 流经其 AI 网关。

上周公司 ShipNYC 大会结束后，我们与 Vercel 首席执行官 Guillermo Rauch 进行了对谈，了解他对当前 AI 形势的看法，以及像 Vercel 这样的平台公司如何与各大实验室竞争。以下是经过轻度编辑的对话实录。

感觉今年社区里有一种不同的能量，试点项目少了，大家更关注如何让产品在实际中良好运作。我相信你们在客户那里已经深有体会，但我好奇 Vercel 内部的这段历程是怎样的。

去年主要是做原型验证。一切皆有可能，释放 Agent，人人都能构建，诸如此类。我们这么做了，并且学到了很多，因为我们公司内部自发开发并部署了数百个 Agent，然后你开始接触到 Agent 在生产环境中面临的现实和挑战。

对我来说最大的收获是那些本垒打式的用例，也就是 Agent 的两大杀手级应用。其一当然是编码 Agent。这推动了全球大量的 Token 使用，但当你生产出如此多的软件时，就需要有个地方来部署它。Agent 的第二个杀手级应用是帮助你运营公司的内部 Agent。这里面临的挑战是：如何安全地访问数据？如何审计 Agent 在做什么？如何获取 Agent 为完成一项任务所经历的所有工具调用和访问控制的完整轨迹？

为了解决这个问题，我们提出了一个名为 Eve 的框架，你可以用自然语言描述 Agent 的指令和技能。另一个工具是 Vercel Sandbox，在那里你可以把 Agent 放进一个小笼子里。它仍然可以自由地发挥其智能，但你可以对它能访问哪些数据以及哪些数据可以离开沙箱施加策略控制。

这能帮助你们避免哪些问题？

对于沙箱而言，最大的优势是数据控制。我一直思考的一个真正的 AI 风险是，当你使用 Devin 或 Cursor 这样的编码 IDE 时，如果设置不当，它们可能会基于你的整个代码库进行训练。我记得曾与空客总裁谈及此事。你们拥有数十年积累的、非常特定的航空航天工程 C++ 代码财富。有人进来安装了错误的开发者工具，然后砰的一声，所有代码都跑到云端进行训练了。

我很想多听听第二个杀手级用例的情况。我们都知道编码 Agent，但实际中企业内部 Agent 是怎样的？

你看，那边坐着一位销售人员（在 Vercel 的办公室里）。她负责存量客户。她的工作是扩大现有账户。对她这样的人来说，瓶颈并不在于她的创造力、智慧或建立关系的能力，而在于数据。“我不明白哪些账户增长更快。给我过去两周增加席位最多的五个账户，这样我就能优先安排工作。”她过去无法提出这个问题。她需要等一个一季度项目完成，开发出新的销售仪表盘才行。

我们 Vercel 多年来一直处于这种瓶颈中，这非常令人沮丧，因为在研发端，我们是世界上行动最快的公司。但在销售引擎方面，即 Salesforce 工程那边，我完全不够格。刚开始我一生中都没打开过 Salesforce。现在我觉得自己确实能够影响整个公司，因为 Eve 既可以用于面向客户的 Agent，也可以用来提升生产力。相同的技术，只不过是通过 API。Agent 正在迫使企业开放，这将产生深远的长期影响。如此多的 SaaS 巨头都是靠着锁住你的数据来构建整个王国，而这与 Agent 不相容。

你如何看待客户与大型 AI 实验室之间的关系变化？

去年有很多人选择单一的实验室合作伙伴——声称他们将全部基于 OpenAI 或 Anthropic 来构建。现在他们说，我明白了这一切是如何运作的——模型、适配层、数据平台、沙箱、网关——每个组件都是即插即用的。你可以用 OpenAI，可以用 Anthropic，也可以用 Gemini。我们看到 Gemini 增长很大，尽管新闻上报道不多，因为现在人们正在为生产环境做优化。现实是，当你在为生产优化时，你就会开始关注性价比，而 Gemini 模型有着极佳的性价比特性。你还会引入开放模型，因此 DeepSeek 和 GLM-5.2 正在起飞。数据不会说谎。

也存在你们与实验室直接竞争的领域，对吧？就在几周前，OpenAI 发布了一套新工具，可以直接发布到网上，无需离开 OpenAI 的封闭环境。

他们托管小型网站是自然而然的一步。这对我们来说是一个很好的契机，因为现在人们会把 ChatGPT 视为一个制作网站的工具。然后如果他们继续向模型询问关于虚拟主机的问题，模型就会推荐我们。但你说得对，随着模型或平台增加更多功能，它们会与现有的基础设施平台直接竞争。

我真心认为，此刻我们正在决定模型和 Agent 究竟要不要捆绑在一起。

你是从一个地方获取全部智能？还是从某个提供商那里获取一个模块、一个库或一个构建块，然后在此基础上进行构建？这更像是软件工程一直以来的方式，也正是我们推向市场的东西。我们将成为这一代的 AWS，所以很明显，我们是在为一个开放协议的世界而战。

Known for its cloud infrastructure that allows developers to deploy agents without managing servers, Vercel has quietly become one of the most central companies in AI software. The company currently sees 6 million deployments a day, half of them triggered by coding agents, and more than 1 trillion tokens flow through the company’s AI gateway daily.

After the company’s ShipNYC conference last week, we sat down with Vercel CEO Guillermo Rauch for his take on this moment in AI, and how platform companies like Vercel end up competing with major labs. Here’s a lightly edited transcript.

It feels like there’s a different energy in the community this year, fewer pilot programs and more focus on how to make things work well in practice. I’m sure you’ve seen that a lot with clients, but I’m curious what that journey has looked like within Vercel.

Last year was about prototyping. The sky’s the limit, unleash the agents, everyone can build, and so on. We did that, and we learned a lot because we had hundreds of agents organically developed and deployed within the company, and then you started getting into the realities of agents in production, and some of the challenges.

The biggest lesson for me was the home-run use cases, the two killer apps of agents. One is the coding agent, of course. That’s driving a lot of the token utilization in the world, but when you produce so much software, you need somewhere to put it. The second killer app of agents is the internal agent that helps you run the company. The challenge there is, how do you securely access data? How do you audit what the agent is doing? How do you get a trail of all of the tool calls and access controls that the agent had to incur in order to get a job done?

To solve that, we came up with this framework called Eve, where you can lay out an agents’ instructions and skills in natural language. And another tool is Vercel Sandbox, where you put the agent in a little cage. It can have the freedom still to express its intelligence, but then you can apply policy on what data it can access and what data can leave the sandbox.

What sort of problems does that help you avoid?

For [the] sandbox, the biggest advantage is data control. A real risk of AI that I always think about is, when you get a coding IDE like Devin or Cursor, if you’re in the wrong setting, they may train on your entire codebase. I remember talking to the president of Airbus about this. You have decades of wealth of very specific C++ code for aerospace engineering. Someone comes in and installs the wrong developer tool and boom, all the code goes out to the cloud for training.

I’m curious to hear more about that second killer use case. We all know about coding agents, but what does an internal corporate agent look like in practice?

So, there’s a sales rep sitting out there [in Vercel’s office]. She works on install base. Her job is to grow existing accounts. The bottleneck for people like her has not been her creativity, intelligence, ability to build relationships, it’s been data. “I don’t understand what accounts are growing faster. Give me the five accounts that have added the most seats in the last two weeks, so that I can prioritize my work.” She couldn’t ask that question in the past. She needed to wait until a Q1 project for a new sales dashboard completed.

We were in that bottleneck for years at Vercel, and it was really frustrating because on the R&D side, we’re the fastest-moving company in the world. But on the sales engine, the Salesforce engineering [side], I was so incompetent. I had never opened Salesforce in my life when I started.

Now I feel like I can actually have impact across the entire company, because Eve can be used for our customer-facing agents and can be used to improve productivity. Same technology, it’s just APIs. Agents are forcing companies to open up, and that will have dramatic long-term implications. So many of these SaaS giants build their entire kingdoms on trapping your data, and that’s incompatible with agents.

How do you see client relationships with the big AI labs changing?

Last year there were a lot of people picking one lab partner — saying they would build everything on OpenAI or Anthropic. Now they’re saying, I understand how this all works — model, harness, data platform, sandbox, gateway — every piece is plug and play. You can use OpenAI, you can use Anthropic, or you can use Gemini. We’re seeing a lot of growth of Gemini, even though it’s not on the news as much, because people are optimizing for production now. The reality is, when you’re optimizing for production, you start looking at a price/performance, and Gemini models have awesome price/performance characteristics. You also bring in open models, so DeepSeek and GLM-5.2 are taking off. The data doesn’t lie.

There are places where you’re in direct competition with the labs too, right? Just the other week, OpenAI released a new set of tools that publish directly to the web without having to leave the OpenAI enclave.

It’s a natural next step for them to host little websites. And it’s a great opening for us, because now people will think of ChatGPT as a tool for making websites. And then if they keep asking the model questions about web hosting, the model recommends us. But you’re right, as the models or platforms add more capabilities, they come in direct competition with the infrastructure platforms that already exist.

I really think at this point we’re deciding on whether the model and the agent are going to be coupled.

Do you get all your intelligence from one place? Or do you get a module or a library or a building block from one provider, and then you build on top of it. That’s more like software engineering has always been, and that’s really what we’re bringing to market. We’re going to be the AWS of this generation, so obviously we’re fighting for a world of open protocols.