请问HN：你们有没有关于AI代理的成功案例？
首先，我想澄清一下我对AI代理的定义：一个利用大型语言模型（LLM）的脚本并不是代理，它只是自动化，或者说是一种“工作流程”（根据Anthropic的说法）。AI代理是一个系统，它通过工具和记忆收集关于输入的信息，从而处理复杂的语义输入，最终生成某种“基于判断”的响应。
例如，编写一首关于你所在城市当前天气的脚本——这属于自动化。一个用于客户支持的聊天机器人——这就是代理。Codex/Claude Code——也是代理。
主要的区别在于，代理会根据需要解决的任务构建自己的上下文。
我听到很多关于AI代理的讨论，但我还没有看到太多成功案例能够很好地契合并提供比开发和维护成本更高的价值（编程代理是个例外）。所以，请分享你们的AI代理成功故事！
查看原文
Firstly to clarify what I mean with AI agent: A script which utilizes an LLM is not an agent. It's automation, alternatively a "workflow" (according to Anthropic). An AI agent is a system which is handling complex semantic input by gathering information about the input via tooling + memory and then lastly process it into some "judgement based" response.<p>The shell script which writes a poem about the current weather in your city -> automation.<p>A chatbot for customer support -> agent. Codex/Claude Code -> agent.<p>Primary difference being that an agent build its own context depending on the task that it needs to solve.<p>I've heard a lot of buzz around AI agents, but I've yet to see that many success stories where they fit nicely and provide more value than what they cost to develop and maintain (exception being coding agents). So, please share your AI agent success stories!