24hrs Inside the $300M Startup Building the Infrastructure Layer for AI Agents

章节 1：AI智能体基础设施的控制之战

📝 本节摘要：

本章介绍Guild AI如何为企业中的AI智能体提供部署、监控与成本控制能力，并以淘金热中的“镐和铲子”作比喻，解释AI基础设施的价值。内容还讨论了智能体从实验走向生产后，如何负责任地使用模型，以及盲目追求Token消耗可能带来的问题。

[原文] [Will Phillips]: Millions of dollars poured into AI, and there's no visibility into how this money is being used.

[译文] [Will Phillips]: 数百万美元涌入AI领域，但人们却无法看清这些钱究竟是怎么被使用的。

[原文] [James Everingham]: Access to these AI tools. It's so exciting. You're able to do so much more than you were before. And spending tokens felt like candy. Over time, you have to think about: are you actually getting your ROI for everything that you're building with your tokens?

[译文] [James Everingham]: 能够使用这些AI工具，真的令人兴奋。你可以完成比过去多得多的事情，花Token的感觉就像吃糖一样。但随着时间推移，你必须想清楚：你用这些Token构建的一切，真的带来了相应的投资回报吗？

[原文] [Will Phillips]: The wave of AI agent adoption has arrived. And agents are now taking real actions inside critical systems. Research predicts 40% of enterprise software to have AI agents built in

[译文] [Will Phillips]: AI智能体的采用浪潮已经到来。如今，智能体正在关键系统中执行真实操作。研究预测，到今年年底，40%的企业软件都将内置AI智能体。

[原文] [Will Phillips]: by the end of this year. But the faster they multiply, the harder it is to track them. But the startup I'm visiting today is building the layer that puts them back in control.

[译文] [Will Phillips]: 但它们增长得越快，就越难以追踪。而我今天要探访的这家初创公司，正在构建一个能让企业重新掌控这些智能体的基础层。

[原文] [James Everingham]: Language modeling really captures a lot of what thinking might be. If you can understand what happens next in the world, it's quite a powerful thing to be able to do.

[译文] [James Everingham]: 语言建模确实捕捉到了许多类似思考的东西。如果你能理解世界接下来会发生什么，那将是一种非常强大的能力。

[原文] [Will Phillips]: Today, I'm going behind the scenes with Guild AI, a startup building the control plane for the AI agents running inside a company. It lets you deploy them, see exactly what they're doing, and easily reduce their token cost across any model provider. The team came out of Meta's developer infrastructure group, and they've raised a combined $44 million led by firms like Google Ventures and Costanoa Ventures.

[译文] [Will Phillips]: 今天，我将深入探访Guild AI。这家初创公司正在为企业内部运行的AI智能体打造控制平面，让企业能够部署智能体、准确了解它们在做什么，并在任何模型供应商的平台上轻松降低Token成本。这个团队来自Meta的开发者基础设施部门，已累计融资4400万美元，领投方包括Google Ventures和Costanoa Ventures等机构。

[原文] [James Everingham]: Coding is not the hard part. It's the tedious part.

[译文] [James Everingham]: 写代码并不是最难的部分，只是最枯燥的部分。

[原文] [Will Phillips]: Last year, running an AI agent was an experiment. But now, they're in full production. And the companies deploying them are waking up to the cost. So, the real question I want to understand is whether the layer that controls these agents ends up mattering as much as the models themselves.

[译文] [Will Phillips]: 去年，运行AI智能体还是一种实验。但现在，它们已经全面投入生产，而部署这些智能体的公司也开始意识到其中的成本。因此，我真正想弄清楚的是：控制这些智能体的基础层，最终是否会和模型本身一样重要。

[原文] [James Everingham]: It's all of our careers right now, and it's all coming to this point.

[译文] [James Everingham]: 这关系到我们所有人的职业生涯，而现在一切都正在走到这个关键节点。

[原文] [James Everingham]: Jim's like, "Wouldn't it be cool to do this for 40 million developers instead of 40,000?" And the answer is yes.

[译文] [James Everingham]: Jim当时说：“如果我们服务的不是4万名开发者，而是4000万名开发者，那不是很酷吗？”答案当然是肯定的。

[原文] [Will Phillips]: Going behind the scenes with a company called Guild AI. Going to catch up with James for a coffee first, the founder. He is an absolute Silicon Valley veteran when it comes to the developer community. He spent a ton of time at Meta. How I would explain Guild really simply is: normal AI tools are like chatbots. You can phone up a friend and get advice. AI agents, which is where they predominantly sit, is engaging someone to do the work for you. They're kind of building the office that all of those agent workers sit in. It allows you to be able to build agents, monitor what's going on, more specifically monitor costs at a very detailed level.

[译文] [Will Phillips]: 今天要深入探访一家叫Guild AI的公司。我会先和创始人James喝杯咖啡聊聊。说到开发者社区，他绝对算得上是硅谷老兵，在Meta工作了很长时间。简单来说，普通AI工具就像聊天机器人，你可以打电话给朋友寻求建议。而AI智能体——这正是他们主要专注的领域——就像是请人替你完成工作。他们正在打造的，有点像是所有这些智能体员工办公的办公室，让你能够构建智能体、监控它们的运行情况，更重要的是，以非常细致的程度监控成本。

[原文] [Will Phillips]: And we're in what I would call this gold rush. The ones who actually win are the ones selling the pickaxes and the shovels. And as Guild AI is an AI infrastructure company that's completely neutral, it allows you to track models on any provider, they are selling the very valuable pickaxes and shovels. As you can imagine, I'm super excited to go behind the scenes with this team, see what's happening in their world, hear more about their philosophy about where this is all actually heading, because if anyone knows, it's these guys.

[译文] [Will Phillips]: 我们正处在我所说的这场淘金热中。真正获胜的，往往是那些卖镐和铲子的人。Guild AI是一家完全中立的AI基础设施公司，可以让你追踪任何供应商提供的模型，所以他们卖的正是非常有价值的“镐和铲子”。可以想见，我非常期待深入了解这个团队，看看他们的世界正在发生什么，听听他们对于这一切最终走向的思考，因为如果有人真正了解答案，那一定就是他们。

[原文] [Will Phillips]: Coffee comes first.

[译文] [Will Phillips]: 先喝咖啡。

[原文] [Will Phillips]: Lovely to meet you.

[译文] [Will Phillips]: 很高兴见到你。

[原文] [James Everingham]: Good to meet you, too.

[译文] [James Everingham]: 我也很高兴见到你。

[原文] [Will Phillips]: Oh, well, thank you.

[译文] [Will Phillips]: 噢，那谢谢你。

[原文] [James Everingham]: You're welcome.

[译文] [James Everingham]: 不客气。

[原文] [Will Phillips]: I had this analogy of the pickaxes and shovels, but in the case of Guild, how does it work?

[译文] [Will Phillips]: 我刚才用了“镐和铲子”这个比喻，但具体到Guild，它是怎么运作的？

[原文] [James Everingham]: Well, as you said, back in the gold rush, a lot of people made money from selling pickaxes and shovels to go find the gold. So, if you're going to carry that analogy forward, like with AI, there are tools that people are building to help create AI. I would think of inference, GPUs, training pipelines, all of this stuff to help make AI better. But then, what happens when you actually get the gold, you know? Like that's sort of the next problem. And that was a whole 'nother industry. That led to jewelry, to circuitry, to all of these things that would apply the gold. So, I would say that metaphor would be kind of what we're going after is, okay, you have AI.

[译文] [James Everingham]: 正如你所说，在淘金热时期，很多人靠卖镐和铲子给淘金者赚钱。如果把这个比喻延伸到AI领域，人们正在打造各种帮助创造AI的工具。我想到的包括推理、GPU、训练流水线，以及所有帮助AI变得更好的东西。但当你真的挖到金子之后，会发生什么呢？这其实是下一个问题，也催生了一个完全不同的行业，带来了珠宝、电路，以及所有能够应用黄金的东西。所以我想说，我们追求的就是这个比喻中的下一步：好，你已经拥有了AI。

[原文] [James Everingham]: Now, how do you fabricate it responsibly and use it in your systems? And that's the problem that we're solving at Guild.

[译文] [James Everingham]: 接下来，你如何负责任地构建它，并把它应用到自己的系统中？这就是Guild正在解决的问题。

[原文] [Will Phillips]: You were about to say something about— I thought we couldn't miss.

[译文] [Will Phillips]: 你刚才似乎正要说些什么——我觉得这个话题不能错过。

[原文] [James Everingham]: There's things that trigger me in the industry, in the way certain companies are moving, and you said two words: token maxing. So, yeah, this was something that we heard recently, where we're judging people on the amount of tokens that they were consuming and saying, like, if you're not consuming a lot of tokens, you're probably not being effective. That might sound reasonable, but let me reframe that to a different industry. Let's say you're a delivery or a freight company and you decide to start measuring the effectiveness of your crew by how much fuel they burn. So, you're going to do fuel maxing. That's probably going to create some strange behaviors, right?

[译文] [James Everingham]: 行业里有些事情会让我很警觉，尤其是某些公司的做法。你刚才提到了两个词：Token最大化。最近我们听到一种说法，说可以根据一个人消耗了多少Token来判断他的工作效果；如果你没有消耗很多Token，那你可能就做得不够有效。这听起来似乎有道理，但让我换一个行业来解释。假设你是一家快递或货运公司，决定根据员工烧掉多少燃油来衡量他们的工作效率。于是你开始追求“燃油最大化”。这大概会导致一些奇怪的行为，对吧？