Ask HN: 代理工作流如何抵消 AI 技术债务？
2 分•作者: l33tbro•大约 2 个月前
我不太清楚该怎么说。但我接手的、需要完成的项目都有一个共同的奇怪特质：它们是“未设计的”。
我认为这可能是因为，之前在基本设计原则和功能方面考虑过的流程和与上下文相关的代码，与其说是缺失，不如说是被模拟了。这给人一种已经被考虑过的印象，这对于项目经理来说很难察觉。
所以，我现在花了很多时间回去设计这些“未设计的”工作。手动重建那些被抽象化或完全遗漏的东西，而项目管理层的人却认为“这不会花你很长时间”。我认为这归因于这种隐蔽的问题，即事物看起来像是被设计过的。
当大规模部署时，这应该如何运作？当人工智能产生如此多的技术债务时，效率的提升真的会开始变成损失吗？
我不确定基于智能体的流程实际上如何解决这个问题。似乎你可以让智能体做一些事情，但它们仍然会犯很多错误，你需要介入并重建。
查看原文
I don't know quite how to put it. But projects I inherit and am supposed to get over the line have this same strange quality: they are 'undesigned'.<p>I believe this may be because processes and context-related code that was previously considered with basic design principles and functionality seem not so much to be missing, but they have they seem to be simulated. Giving the impression of being considered, which is quite difficult for a project manager to catch.<p>So I'm now spending a lot more time now going back and designing this 'undesigned' work. Rebuilding things manually that was abstracted past or missed entirley, which people at the project management level think "won't take you that long". Which I attribute to this insidious problem of things looking like they were designed.<p>How is this meant to work when deployed at scale? When there is so much technical debt generated by AI, do the efficiency gains actually start to be losses?<p>I'm not sure how agentic workflows can actually solve this. It seems as if you can have agents doing stuff, but they're still going to get a lot wrong and you'll need to drop someone in and rebuild.