Vercel CEO reveals his Agentic Engineering Workflow

章节 1：安全债务与后AGI时代的控制权

📝 本节摘要：

Guillermo Rauch指出，随着攻击面扩大，理解新软件带来的安全债务是巨大挑战。他认为后AGI时代的关键在于人类保持对创意、网络安全及智能体运行的控制权。尽管农业曾占人口90%，如今不足1%，但软件开发已成为人类的创意媒介，而非单纯劳动。AI正在赋能更宏大的叙事与世界观构建，如《指环王》重制，激发了人们构建新世界的热情。

[原文] [Guillermo Rauch]: One of my biggest challenges today in adopting new software, both personally and as Vercel, is understanding the security debt that I'm taking on. Just the surface of attack has become so massive as of recently. We probably need to start writing code in different programming languages. The challenge that I think a lot of organizations have today is that the amount of latent vulnerabilities of all of this human-written code is almost immeasurable. We could eradicate supply chain attacks if we moved more and more of this development life cycle to sandboxes.

[译文] [Guillermo Rauch]: 如今我在采用新软件时面临的最大挑战之一，无论是个人层面还是作为 Vercel，就是理解我所承担的安全债务。最近攻击面变得极其庞大。我们可能需要开始用不同的编程语言来写代码。我认为许多组织今天面临的挑战是，所有人类编写的代码中潜在漏洞的数量几乎是无法衡量的。如果我们把开发生命周期的更多部分转移到沙箱中，我们就能根除供应链攻击。

[原文] [David Ondrej]: All right. So, Gilmo, what will matter in the post-AGI era?

[译文] [David Ondrej]: 好的。那么，Gilmo，在后通用人工智能（AGI）时代，什么将变得重要？

[原文] [Guillermo Rauch]: I think our ability to stay in control,

[译文] [Guillermo Rauch]: 我认为我们保持控制权的能力，

[原文] [David Ondrej]: okay,

[译文] [David Ondrej]: 好的，

[原文] [Guillermo Rauch]: as humans. And by that I mean creative control, control over the risks on cyber security.

[译文] [Guillermo Rauch]: 作为人类。我的意思是创意控制权，以及对网络安全风险的控制。

[原文] [David Ondrej]: Yeah.

[译文] [David Ondrej]: 是的。

[原文] [Guillermo Rauch]: Control over how we make the agents work for us ultimately. You know, I've been speaking a lot about the leverage is shifting one layer of abstraction higher into sort of the hard engineering, the software factory, the fleets of agents that you orchestrate. I'm a huge believer in our taste and our creativity as human beings because we're usually appealing to other human beings and we love to buy the stories and products and services from other human beings. You know, we like to think that we make perfectly rational decisions as humans, but a lot of the time we're actually buying into the stories and the vibes and the brands and the flavors and the tastes. Ultimately, agents are going to be hopefully an enabler of greater and greater storytelling and greater and greater human ambition. I like what you said about the higher level work because

[译文] [Guillermo Rauch]: 最终是对我们如何让智能体为我们工作的控制。你知道，我谈了很多关于杠杆正在向更高一层抽象转移，转向所谓的硬工程、软件工厂，以及你编排的智能体集群。我坚信我们作为人类的品味和创造力，因为我们通常是在吸引其他人类，我们喜欢从其他人类那里购买故事、产品和服务。你知道，我们喜欢认为作为人类我们做出完全理性的决定，但很多时候我们实际上是在为故事、氛围、品牌、风味和品味买单。最终，智能体有望成为越来越强大的叙事和越来越宏大的人类抱负的赋能者。我喜欢你所说的高层次工作，因为

[原文] [David Ondrej]: people always are afraid of like this job going away, this task going away and you know farming was something that like 90% of population did 200,000 years ago, 200 years ago

[译文] [David Ondrej]: 人们总是害怕这份工作消失，这项任务消失。你知道，农业是20万年前、200年前90%的人口从事的事情，

[原文] [David Ondrej]: and now nobody that works in farming, less than 1% of people. So thinking of this is like filling out a form just feels like manual labor, you know. So what will that look like? Will it be like managing fleets of agents like deciding which models, deciding like where they run in terms of the hardware? It's so funny because one of my best performing tweets recently was I do this for fun. You know, I build software to relax.

[译文] [David Ondrej]: 而现在从事农业的人不到1%。所以把填表看作只是体力劳动，你知道。那会是什么样子？会是管理智能体集群吗？比如决定使用哪些模型，决定它们在硬件上运行的位置？这很有趣，因为我最近表现最好的一条推文是“我做这个是为了乐趣”。你知道，我构建软件是为了放松。

[原文] [Guillermo Rauch]: Yeah.

[译文] [Guillermo Rauch]: 是的。

[原文] [Guillermo Rauch]: I build software for work and then I go home and I build more software. I build software with my kids. I build software with my wife. Like I'm always building software. So that's where I think the analogy breaks down with farming. And by the way, like there are some people that are legitimately love farming and maybe they, you know, over hundreds of years, millennia, they'll still do it.

[译文] [Guillermo Rauch]: 我为了工作构建软件，然后回家继续构建更多软件。我和孩子们一起构建软件，我和妻子一起构建软件。我总是在构建软件。所以我认为这与农业的类比在这里失效了。顺便说一下，确实有些人真心热爱农业，也许在几百年、几千年后，他们仍然会做这件事。

[原文] [David Ondrej]: It's called gardening.

[译文] [David Ondrej]: 那叫园艺。

[原文] [Guillermo Rauch]: Yeah. And they keep their little like mini farm in their balcony and that kind of stuff. But I think a lot of us do this because we this is our medium for creativity and there's no taking that away. There's not putting that genie back in the bottle as far as I'm concerned. But it does mean that I mean you kind of see this with models these days being benchmarked over their ability to create new worlds. There was that Karpathy discontinuity tweet of, you know, first he introduced vibe coding, then he introduced, oh, by the way, the models are getting so good that they can create entire worlds and they can re-render the Hobbit or Lord of the Rings.

[译文] [Guillermo Rauch]: 是的。他们在阳台上种着小小的迷你农场之类的东西。但我想我们很多人做这件事是因为这是我们的创意媒介，没有人能夺走它。就我而言，那个精灵不会被放回瓶子里。但这确实意味着，你看，如今模型正在根据其创造新世界的能力进行基准测试。有 Karpathy 的那条关于“不连续性”的推文，你知道，他首先引入了“氛围编程”（vibe coding），然后他说，哦，顺便说一下，模型变得如此好，以至于它们可以创造整个世界，可以重新渲染《霍比特人》或《指环王》。

[原文] [Guillermo Rauch]: And so I do think that we're getting into the next level of how much free range of creativity is enabled by AI, creating new worlds, creating games. I've spoken of how the web itself, the web that we will consume as humans is becoming like ultra fancy. Like the experiences that stand out to you are the ones that just like really go out of their way in terms of wowing you. 3D and animations and particle effects and things like that. And so it's amazing that like literally we can do anything now. We can build anything. And people are flexing that muscle. Like I see a lot of people that are building stuff that they're just building it because they couldn't before and they're deriving a lot of joy from building which I think is really good.

[译文] [Guillermo Rauch]: 所以我认为我们正在进入下一个层级，即AI如何赋能更大范围的创造力，创造新世界，创造游戏。我谈过Web本身，我们作为人类消费的Web正在变得极其华丽。那些让你印象深刻的体验，是那些真正在让你惊叹方面不遗余力的体验。3D、动画和粒子效果等等。所以令人惊叹的是，我们现在真的可以做任何事。我们可以构建任何东西。人们正在锻炼这块肌肉。我看到很多人在构建东西，他们只是构建，因为以前他们做不到，并且从构建中获得了很大的乐趣，我认为这非常好。

[原文] [David Ondrej]: I've been coding with AI for nearly four years. In 2023 I was the first to start the build anything trend and since then a lot has changed. We have different models, different tools, different agents. But despite all of that change, every single project I've built has one thing in common. I always use Supabase as my back end. I'm not even thinking about what DB do I use. I just use Supabase. It's simple, it's scalable, and it's agent first. In fact, it has never been easier to get started. Now, the AI agents can literally build the entire back end for you because Supabase now works inside of Codex and Claude Code. We all know that with AI, you can write code faster than ever. But where people get stuck is the back end. The database, the login, the storage, all the stuff you need to get right. And

[译文] [David Ondrej]: 我用AI编程快四年了。2023年，我是第一个开启“构建任何东西”趋势的人，自那以后变化很大。我们有不同的模型、不同的工具、不同的智能体。但尽管有所有这些变化，我构建的每一个项目都有一个共同点。我总是使用 Supabase 作为后端。我甚至不思考用什么数据库。我就用 Supabase。它简单、可扩展，并且以智能体为先。事实上，开始使用它从未如此容易。现在，AI智能体真的可以为你构建整个后端，因为 Supabase 现在可以在 Codex 和 Claude Code 内部工作。我们都知道，借助AI，你可以比以往更快地编写代码。但人们卡住的地方是后端。数据库、登录、存储，所有你需要弄对的东西。而