“带工具的代理”是唯一有效的 LLM 应用吗？
2 分•作者: gmarkwa•大约 1 个月前
Brex 的 CEO 曾说过这样的话，我理解他的出发点，但在基于 LLM 的应用程序中，端到端工作流是否有用武之地？归根结底，我同意最先进的软件应该完全由代理（agents）使用，但我无法确定每个产品是否都应该是一个代理。我个人非常讨厌大多数代理式应用程序的用户体验。我有一个关于未来产品的幻想，即我可以简单地输入 X 并获得输出 Y，而无需对每一个小细节进行反馈。当然，在设计产品时，它需要一个过程，但如果我能告诉我的手机“买晚餐”，而它就能完成，而无需在每一步都打扰我呢？如果它足够好呢？我认为，只要应用程序不确定输出的准确性，可对话的代理就是必要的，但我怀疑 LLM 应用程序的用户体验是否可以更像一个富人修理汽车，即“这是我的情况，你做好你需要做的事情，别来烦我。”
查看原文
Brex's CEO said this and I understand where he comes from, but is there a place for end to end workflows in LLM based applications? At the end of the day I agree that cutting edge software should be fully usable by agents, but I can't decide if every product should be an agent. I personally hate the UX of most agentic applications. I have this fantasy of future products where I can simply enter input X and get output Y without having to give feedback on every little thing. Of course when you're designing something it needs to be a process, but what if I can tell my phone to "buy dinner" and it just does it without bothering me at every step? What if it was just good enough? I think a conversable agent is necessary as long as the application is unsure of the accuracy of the output, but I wonder if the UX of LLM applications could be more like a rich person fixing their car, like here is my situation, just do what you need to do and don't bother me.