我们在 Vercel 构建代理时学到的东西

Nov 6, 2025

2025年11月6日

Agents present incredible promise for increased productivity and higher quality outcomes in enterprises. Companies are already using them to streamline customer support, code reviews, and sales operations.

代理在企业中展现出提高生产力和更高质量成果的巨大潜力。公司已经在使用它们来简化客户支持、代码审查和销售操作。

When building custom internal agents, the challenge isn't whether AI can create value, it's identifying the problems it's ready to solve today, at a cost that makes sense for the business.

在构建自定义内部代理时，挑战不在于AI是否能创造价值，而在于识别它今天准备解决的问题，以及以对业务有意义的成本。

At Vercel, we are going through the same AI transformation as our customers. We use our own products to build agents that help us move faster and spend more time on meaningful work.

在 Vercel，我们正在经历与客户相同的 AI 转型。我们使用自己的产品构建代理，帮助我们更快地行动，并花更多时间在有意义的工作上。

After months of experimentation, we’ve turned our learnings into a repeatable methodology for finding and investing in AI projects that have the highest likelihood of creating significant business impact.

经过几个月的实验，我们将所学转化为一种可重复的方法论，以寻找和投资于那些最有可能产生重大商业影响的AI项目。

Over time AI will touch nearly every workflow, handling complex tasks like our own code review and anomaly investigation agent. Our intuition for what agents can do is skewed towards high expectations because coding agents like these are so amazing.

随着时间的推移，AI将触及几乎每个工作流程，处理复杂任务，例如我们自己的 代码审查和异常调查代理。我们对代理能做什么的直觉倾向于高期望，因为像这样的编码代理实在太棒了。

But most companies don’t have the engineering capacity to productionize that level of internal use case, and today’s models still face limits in reliability and precision in other domains. This is why we need to select problems that fit what today's frontier models are well suited for.

但大多数公司没有工程能力将这种内部用例投入生产，而今天的模型在其他领域仍面临可靠性和精确度的限制。这就是为什么我们需要选择适合今天前沿模型的解决问题的方法。

We've learned that the highest likelihood of success for current-generation agentic AI comes from work that requires low cognitive load and high repetition from humans.

我们了解到，当前一代代理AI成功的最高可能性来自于需要低认知负荷和高重复性的工作。

The sweet spot is human work with low cognitive load and high repetition

最佳点是低认知负荷和高重复性的人工工作

These ta...