A Billion Token Lesson: Because You Can ≠ You Should

A Billion Token Lesson: Because You Can ≠ You Should 十亿 Token 的教训：能做 ≠ 该做

Agent Autopsy, Day 6. Last week I spent a weekend building a product nobody asked for. I didn’t catch it in time. Let me tell you how it happened. 智能体剖析，第 6 天。上周末，我花了一整天时间开发了一个根本没人需要的产品。我没能及时意识到这一点。让我来告诉你这是怎么发生的。

The trap 陷阱

Nicolas Fränkel published a piece on designing teams of AI agents — specialist subagents that collaborate on software engineering tasks. Planner, challenger, coder, tester. Each with a focused system prompt, each with limited tools, orchestrated through a skill. It’s good work. Nicolas Fränkel 发表了一篇关于设计 AI 智能体团队的文章——即协作完成软件工程任务的专业子智能体。规划者、挑战者、编码者、测试者。每个智能体都有专注的系统提示词，拥有有限的工具，并通过某种技能进行编排。这是一项出色的工作。

I read it and immediately thought: we should build this for SPFx. A TypeScript specialist harness. Four agents — Architect, Scaffolder, Builder, Verifier — in a deterministic pipeline. Zod-validated handoffs between them. Compliance injected at build time, not bolted on after. Terminal CLI, BYO token, no desktop app. I had the architecture designed in 15 minutes. The pipeline spec. The agent system prompts. The handoff packet schema. Everything. 我读完后立刻想到：我们应该为 SPFx 构建一套这样的系统。一个 TypeScript 专家框架。四个智能体——架构师、脚手架构建者、构建者、验证者——在一个确定性的流水线中运行。它们之间通过 Zod 验证进行任务交接。合规性在构建时注入，而不是事后修补。终端 CLI，自带 Token，无需桌面应用。我用了 15 分钟就设计好了架构。流水线规范、智能体系统提示词、交接数据包模式。一切就绪。

Then I stopped and asked: is anyone actually looking for this? The answer was no. 然后我停下来问自己：真的有人需要这个吗？答案是没有。

The numbers don’t lie 数据不会撒谎

SPFx developers are a niche. SharePoint is a niche. Agent-orchestrated SPFx development is a niche of a niche — maybe a few hundred people worldwide who’d even understand the value proposition. Of those, how many would switch from Copilot in VS Code to a terminal-based agent pipeline? Zero. The market had already spoken. It was speaking before I even asked the question. But the exciting part isn’t the market analysis. It’s what I did next. SPFx 开发者是一个小众群体。SharePoint 是一个小众领域。由智能体编排的 SPFx 开发更是小众中的小众——全球可能只有几百人能理解其价值主张。在这些人中，有多少人会放弃 VS Code 中的 Copilot 而转向基于终端的智能体流水线？零。市场已经给出了答案。甚至在我提出这个问题之前，市场就已经在回答了。但最令人兴奋的不是市场分析，而是我接下来的行动。

What I did instead 我转而做了什么

I binned the harness. No commit. No repo. One weekend I’m not getting back. Then I published the things people were actually asking for: A learn hub with 10 methodology patterns — 85,000 words of production agent knowledge, free, no email gate. Live at workswithagents.dev/learn. 我扔掉了那个框架。没有提交代码，没有创建仓库。那个周末我回不来了。然后，我发布了人们真正需要的东西：一个包含 10 种方法论模式的学习中心——85,000 字的生产级智能体知识，免费，无需邮箱注册。网址：workswithagents.dev/learn。

A benchmark of 12 local LLMs on real agent coding tasks. SmolLM3-3B at 93.3% — beating Claude Sonnet 4. Qwen2.5-1.5B at 85% on 940MB. DeepSeek-R1 collapsed to 27.5% because reasoning training is toxic below 3B. Data nobody else has published. The harness took a weekend. The learn hub and benchmark took a few hours. One made 0 people’s lives better. The other gave people answers they were actually searching for. 我还发布了一份针对 12 个本地大模型的真实智能体编码任务基准测试。SmolLM3-3B 达到了 93.3% 的准确率——击败了 Claude Sonnet 4。Qwen2.5-1.5B 在 940MB 的体积下达到了 85%。DeepSeek-R1 则跌至 27.5%，因为推理训练在 3B 以下的模型中表现不佳。这些数据是其他人从未发布过的。那个框架花了我一个周末，而学习中心和基准测试只花了几个小时。前者没有改善任何人的生活，而后者给了人们他们真正寻找的答案。

The billion token lesson 十亿 Token 的教训

Agents make us dangerous. Not in the Skynet way — in the I can build anything in an afternoon way. When an agent can scaffold a project in 30 seconds, write a pipeline in 2 minutes, and deploy to production before lunch, the cost of building drops to nearly zero. That sounds great. It’s actually a trap. 智能体让我们变得“危险”。不是像天网那样，而是指那种“我一下午就能造出任何东西”的能力。当智能体能在 30 秒内搭建项目、2 分钟内写好流水线、午饭前完成生产部署时，构建成本几乎降为零。这听起来很棒，但实际上是个陷阱。

Without the friction of effort, there’s no natural filter against bad ideas. In the pre-agent world, you wouldn’t build an SPFx specialist harness because the thought of spending 40 hours on it was exhausting. The idea died on the whiteboard where it belonged. Now an agent can build it for you in 4 hours. The only gate left is your judgment. 没有了努力带来的阻力，也就没有了过滤糟糕想法的天然屏障。在智能体时代之前，你不会去构建一个 SPFx 专家框架，因为想到要花 40 个小时就让人精疲力竭。那个想法会死在白板上，那才是它该待的地方。现在，智能体可以在 4 小时内为你完成它。剩下的唯一关卡就是你的判断力。

That’s the billion token lesson. It’s not about saving compute. It’s about saving the thing you can’t get back: your weekend, your focus, your credibility. 这就是十亿 Token 的教训。重点不在于节省计算资源，而在于节省那些你无法挽回的东西：你的周末、你的专注力、你的信誉。

The rule 准则

Before you tell an agent to build something, ask: Is anyone looking for this? Not “could someone use this?” Not “is this technically possible?” Not “would it be cool?” Is there someone, right now, who would pay money or attention for this thing to exist? If the answer is no — close the terminal. Go work on something people actually want. 在让智能体构建任何东西之前，请问自己：有人在寻找这个吗？不是“有人能用吗？”，不是“技术上可行吗？”，也不是“这看起来酷吗？”。而是：现在是否有人愿意为这个东西的存在而付费或投入关注？如果答案是否定的——关掉终端。去做人们真正想要的东西。

“Because we can” is never the right reason. The market doesn’t care what you’re capable of building. It cares what you actually shipped that someone needed. The SPFx harness taught me this the hard way — with a full weekend I’m not getting back. Cheapest lesson I ever learned. “因为我们能做”永远不是正当理由。市场不在乎你有能力构建什么，它只在乎你交付了什么别人真正需要的东西。SPFx 框架用惨痛的方式教会了我这一点——我损失了一个周末。这是我学到过最“昂贵”的教训。