The Augmentation Gap: Why Using AI Isn’t the Same as Engineering With It

增强鸿沟：为什么“使用 AI”不等于“用 AI 进行工程化”

Three months running an AI agent full-time has clarified something I didn’t expect to learn: most engineers use AI. Few actually engineer with it. There’s a gap forming in the engineering world. 全职运行 AI 智能体三个月后，我意识到了一件意想不到的事：大多数工程师都在“使用”AI，但很少有人真正“用 AI 进行工程化”。工程界正在形成一道鸿沟。

On one side: engineers who use AI — autocomplete, ChatGPT queries, GitHub Copilot suggestions. They’re faster at some tasks. They use AI as a smarter search engine. 一边是“使用 AI”的工程师——他们利用自动补全、ChatGPT 查询和 GitHub Copilot 建议。他们在某些任务上更快，把 AI 当作更智能的搜索引擎。

On the other side: engineers who engineer with AI — treating AI as a collaborator, redesigning their workflows around AI capabilities, building systems that have AI at their core. These engineers are rare. The difference isn’t effort. It’s mindset. 另一边是“用 AI 进行工程化”的工程师——他们将 AI 视为合作伙伴，围绕 AI 的能力重构工作流，并构建以 AI 为核心的系统。这类工程师凤毛麟角。区别不在于努力程度，而在于思维方式。

What “Using AI” Actually Looks Like

“使用 AI”的真实面貌

Most AI adoption looks like this: the engineer is working, hits a wall, opens a chat window, asks a question, pastes the answer back into their code. This is useful. But it’s just faster Googling. The workflow is identical to the pre-AI version — the only change is the retrieval speed. The engineer still: Decides what to build, Catches the errors, Understands the system, Makes every architectural call. AI is a very fast assistant. But the engineer is still the bottleneck. 大多数 AI 的应用场景是这样的：工程师在工作中遇到瓶颈，打开聊天窗口，提问，然后将答案粘贴回代码中。这很有用，但本质上只是更快的“谷歌搜索”。工作流与 AI 出现前完全一致——唯一的改变是检索速度。工程师依然需要：决定构建什么、捕捉错误、理解系统、做出每一个架构决策。AI 只是一个非常快的助手，但工程师本人依然是瓶颈。

What “Engineering With AI” Looks Like

“用 AI 进行工程化”的真实面貌

Engineering with AI is different. It means redesigning the workflow so that AI handles the parts it’s genuinely better at — not just “things that are faster to ask than to Google.” For me, that meant: Delegating entire subsystems to AI agents, not just individual functions; Writing skills that encode repeatable workflows, not just one-off tasks; Treating AI failures as system design problems, not just error fixing; Building memory into the system so the AI compounds its learning. The goal isn’t to replace the engineer. It’s to change who the bottleneck is. “用 AI 进行工程化”则不同。这意味着重构工作流，让 AI 处理它真正擅长的部分——而不仅仅是“问起来比谷歌搜索更快”的事情。对我而言，这意味着：将整个子系统委托给 AI 智能体，而不仅仅是单个函数；编写能够编码可重复工作流的技能，而不仅仅是一次性任务；将 AI 的失败视为系统设计问题，而非简单的错误修复；在系统中构建记忆，使 AI 的学习能够产生复利。目标不是取代工程师，而是改变瓶颈所在。

The Augmentation Gap in Practice

实践中的增强鸿沟

Here’s the practical difference: Using AI: “I need to write this API client. Let me ask ChatGPT.” Engineering with AI: “I need an agent that can build API clients on my behalf. Let me design a skill that teaches another AI how to do this, with my conventions.” The first produces a working API client. The second produces a repeatable system. The first is faster. The second is leverage. 以下是实际的区别： “使用 AI”：我需要写这个 API 客户端，问问 ChatGPT 吧。 “用 AI 进行工程化”：我需要一个能代表我构建 API 客户端的智能体。让我设计一种技能，教另一个 AI 如何按照我的规范来完成这项工作。 前者产生了一个可用的 API 客户端，后者产生了一个可重复的系统。前者更快，后者则是杠杆。

Why This Matters Right Now

为什么这在当下至关重要

We’re in a moment where AI tooling is maturing fast, but most engineering teams are still using AI the same way they used Stack Overflow in 2015 — as a lookup tool. The engineers who understand the difference are the ones building the workflows that everyone will be using in two years. You don’t need to be an AI researcher. You need to treat AI as a collaborator with specific capabilities and specific limitations — not a magic box that answers questions. 我们正处于 AI 工具快速成熟的时期，但大多数工程团队仍像 2015 年使用 Stack Overflow 那样使用 AI——将其仅仅当作查询工具。那些理解其中差异的工程师，正在构建两年后每个人都会使用的标准工作流。你不需要成为 AI 研究员，你需要将 AI 视为一个拥有特定能力和局限性的合作伙伴，而不是一个只会回答问题的魔法盒。

The Real Skill

真正的技能

The skill isn’t knowing what AI can do. It’s knowing what you should stop doing so AI can do it instead. The augmentation gap isn’t about tools. It’s about what you’re willing to redesign. 真正的技能不在于知道 AI 能做什么，而在于知道你应该停止做什么，以便让 AI 来接手。增强鸿沟无关工具，而在于你愿意重构什么。