我如何真正实现自动化：不止于 AI
Blitz.gg 工程主管分享了他长期的自动化经验，强调自动化并非简单的将任务委托给不可控的 AI 模型，而是建立可靠的、可观察的流程。
他认为目前流行的“AI 代理”模式存在信任问题，因为难以保证其决策的正确性。
文章指出，真正的自动化应该像他在 Oracle 的早期经历一样，在“自动化或灭亡”的环境下，直接解决问题，而非依赖不可预测的 LLM。
他批评了开放式访问权限的 “OpenClaw” 风格的自动化方法，认为其缺乏可控性和安全性。
查看原文开头（英文 · 仅前 3 段）
8 min read3 days ago--As the head of engineering for blitz.gg, dyno.gg, and probot.io I do one thing well: I automate my job, all day, everyday and without OpenClaw.The fact is, I’ve been automating things since before “agentic” was a buzzword. That is because when I started my career in the release engineering team at Oracle, it was “automate or die” at every turn. When your pipeline involves 647 manual steps and need to run thousands of tests every night, you don’t philosophize about automation. You just do it.So when the hype machine started screaming about AI agents taking over the world, I had mixed feelings. Not skepticism about AI itself , I think LLMs are genuinely useful. But the framing bothered me.The idea that you should hand the wheel to a model and let it drive. That’s not automation. That’s delegation to something you can’t fully control, observe, or reproduce.The ProblemYes, LLMs can reason about things. They can summarize, classify, and synthesize in ways that would have required entire teams a decade ago. But they’re also non-deterministic, occasionally sycophantic, and depending on which provider you’re using, one deprecation away from your entire workflow collapsing.Honestly the OpenClaw style of “just give the agent access to everything” approach has a few problems:Trust. If an LLM is making decisions autonomously, how do you know it’s making the right ones? You often don’t, until it’s…
※ 出于版权考虑，仅引用前 3 段。完整内容请阅读原文。