判断与想法Notes

我不太关心 Agent 一次能生成多少文件，更关心它做错时能不能停、能不能回滚、下一个人能不能理解发生了什么。I care less about how many files an agent can generate in one pass than about whether it can stop when it is wrong, roll back, and let the next person understand what happened.

真正的工程价值，是把概率性的生成能力装进可观察、可治理的交付系统。The real engineering value lies in fitting probabilistic generation into an observable, governable delivery system.

我更愿意明确写出没有 CI、没有数据集、只支持单机，或者尚未做过真实场景验证。I would rather state plainly that something has no CI, no dataset, runs on a single machine only, or has never been tested in a real setting.

限制不是减分项。对技术工作来说，知道结论在哪些条件下不成立，本身就是能力。Limitations are not a deduction. For technical work, knowing under what conditions a conclusion fails is itself a capability.

“模型会不会被注入”很重要，但更重要的问题是：被注入以后，它到底拿得到什么、改得了什么。Whether a model can be injected matters, but the more important question is what it can actually reach and change once it has been.

因此我会把能力装配、审批、输出过滤、备份恢复看成同一套 Agent 安全边界。So I treat capability assembly, approval, output filtering, and backup and recovery as one agent security boundary.

我更喜欢把复杂工程问题翻译成用户能理解、能选择、能亲手完成的小步骤。I prefer translating complex engineering problems into small steps users can understand, choose between, and complete themselves.

工具可以给建议和护栏，但关键判断应该逐渐回到使用者自己手里。A tool can offer suggestions and guardrails, but the key judgments should gradually return to the person using it.