🔶 温室与透镜：Agentic AI 工作的两种模式

温室和透镜都依靠阳光运作，却以截然相反的方式处理它。温室会捕获环境热量并将其留存，因此其中的一切都会比在室外生长得稍快一些：花朵、香草、杂草，以及土壤里碰巧有什么。透镜则将同样的光线聚焦于单一点，直到那个点燃烧起来。一个汇聚能量，让所有东西同时生长——分散而略显混乱；另一个则是你刻意设计的离散事件：瞄准、聚焦、保持稳定、点燃。

起初，我认为这两种模式可以清晰地映射到软件工程经验上：初级工程师生活在温室中，不加区分地培育杂草和花朵；高级工程师则一次将透镜对准一个定义明确的点。但当我把这个隐喻应用到自己的工作时，这种映射就失效了。两者没有哪个更高级。它们是服务于不同工作的不同工具，而我所看到的许多围绕 Agentic AI 的挫败感，都源于在工作需要一种模式时，人们却运行了另一种模式。

温室模式适合你尚不知道自己想要什么的时候。我的 AxiDraw 绘图仪实验就是纯粹的温室模式：我让 Claude Code 生成了数十个用于制作可绘制艺术作品的工具。大多数结果平平无奇，少数令人愉悦，但没有一个是真正经过规划的。我只是不断尝试，并在事后评估结果。

我制作的一个小型 macOS/SwiftUI 应用也是如此（准确地说，是 Claude 制作的）：它能将 Markdown 转换为 Word 文档。它从来没有规格说明、设计或计划。我让 Claude 构建这个工具，而它的受众只有一个人，生命周期可能也就一周。

这些工具得以出现，是因为尝试新事物的成本已降至近乎为零。在我谈到Claude Code 极其擅长构建 MVP时，我写过这种模式的一个版本：先做一个工具或玩具，看看它是否有用或有趣，然后再决定下一步。

透镜模式则是相反的纪律。你完全清楚“完成”是什么样子，并将一切集中于抵达那里。Agent 提出十件事，而你舍弃其中九件，因为它们不在当前位置与完成状态之间的那条线上。人们谈论 Agentic AI 工具惊人的能力时，通常指的就是这种模式；但值得准确说明这种能力是什么：它是针对已知目标的杠杆。无论是透镜，还是世界上最强大的生成式 AI 模型，当你不知道该把它指向何处时，都可能毫无用处。选择目标，并专注于它，是你的工作。

此外，我认为真正高级的技能，是知道眼前的工作需要哪种模式，识别自己当前正处于哪种模式，并理解当偏移发生时，如何让自己切换到正确的模式。

例如，我坐下来准备交付一个具体功能——显然这是透镜型工作——可四十分钟后，Claude Code 正在搭建一个我从未要求过的管理后台。而我还会点头认同，因为看着东西不断生长，会让人感觉自己在取得进展。反方向的情况也会发生：我本来打算探索，最终却将一个半成形的想法打磨到生产级质量，因为完成一件事令人舒适，而漫游探索则不是。两种情况下，失败都是一样的：我失去了对自己所处模式的判断。

今年 5 月，我曾提出，领域专业知识始终是真正的护城河：在 Agentic 世界中，稀缺资产是一个经过验证的领域心智模型，像一个能够分辨正确答案与看似合理的错误答案的神谕。这个神谕在这里同样适用。在温室中，神谕会从一千株刚冒出的植物中筛选出三株值得保留的；使用透镜时，它会拒绝那些看似合理却错误的输出，因为那些输出会把你带离当前位置与完成状态之间的路径。而发现偏移的也是这个神谕：只有当你把目标记在脑中时，才能察觉自己正在偏离目标。

因此，随着工作推进、你开始进入心流状态时，不妨做一次直觉检查：温室，还是透镜？在合适的情境下，它们都是正当且高效的方法；但专业能力意味着培养出判断何时该使用哪一种的感觉。

A greenhouse and a lens both run on sunlight, and they do opposite things with it. The greenhouse traps ambient heat and holds it, so everything inside grows a little faster than it would outside: flowers, herbs, weeds, whatever happens to be in the soil. The lens takes the same light and concentrates it on a single point until that point catches fire. One gathers energy and lets everything grow at once, diffuse and a little chaotic. The other is a discrete event you engineer: aim, focus, hold steady, ignition.

At first, I thought the two modes mapped cleanly onto software engineering experience: the junior engineer lives in the greenhouse, growing weeds and flowers indiscriminately, while the senior engineer aims a lens at one well-defined spot at a time. But when I applied the metaphor to my own work, the mapping fell apart. Neither mode is the advanced one. They’re distinct tools for distinct jobs, and much of the frustration I see with agentic AI comes from running one mode when the work demands the other.

Greenhouse mode is for when you don’t yet know what you want. My AxiDraw plotter experiments are pure greenhouse: I’ve had Claude Code generate dozens of tools for making plottable art. Most of the results were mediocre, a few were delightful, but none of them were really planned. I just played and evaluated the results afterwards.

The same was true for a small macOS/SwiftUI app I made (well, Claude made) that converts Markdown to Word documents. There was never a spec, a design, or a plan. I asked Claude to build a tool that had an audience of one and a shelf life of maybe a week.

These tools came about because the cost of trying things out has dropped to nearly zero. I wrote about a version of this when I said Claude Code is unreasonably good at building MVPs: build a tool or a toy, see if it’s useful or fun, and go from there.

Lens mode is the opposite discipline. You know exactly what done looks like, and you concentrate everything on getting there. The agent proposes ten things and you discard nine because they aren’t on the line between here and done. When people talk about the shocking power of agentic AI tools, this is usually the mode they mean, but it’s worth being precise about what that power is: leverage on a known target. Both a lens and the most powerful generative AI model in the world can be useless when you don’t know where to point it. Picking the target—and focusing on it—is your job.

Furthermore, I’d argue that the really advanced skill is knowing which mode the work in front of you requires, recognizing which mode you’re currently in, and understanding how to shift yourself to the correct mode when drift has occurred.

For instance, I sit down to ship a specific feature, obvious lens work, and forty minutes later Claude Code is scaffolding an admin panel I never asked for. And I nod along, because watching things grow feels like progress. It runs the other direction too: I set out to explore and end up polishing one half-formed idea to production quality, because finishing is comfortable and wandering isn’t. In both cases the failure is the same: I lost track of which mode I was in.

In May I argued that domain expertise has always been the real moat: the scarce asset in an agentic world is a verified mental model of the domain, like an oracle that tells a right answer from a plausible wrong one. That same oracle works here, too. In the greenhouse, the oracle sifts through the thousand sprouting plants to find the three worth keeping. With a lens, it rejects the plausible-but-wrong output that would pull you off the line between here and done. And it’s the oracle that notices drift: you can only feel yourself wandering off target if you hold the target in your head.

So, as your work progresses and you start to get into your flow, do a gut-check: greenhouse or lens? They’re both legitimate and productive approaches in the right circumstance, but expertise means cultivating the sense of which to apply and when.