<rss xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title>FullStack-Agent - 标签 - lilfry's library</title><link>https://lilfry09.github.io/tags/fullstack-agent/</link><description>FullStack-Agent - 标签 - lilfry's library</description><generator>Hugo -- gohugo.io</generator><language>zh-CN</language><managingEditor>lilfry@sjtu.edu.cn (lilfry)</managingEditor><webMaster>lilfry@sjtu.edu.cn (lilfry)</webMaster><copyright>All rights reserved.</copyright><lastBuildDate>Thu, 06 Aug 2026 20:00:00 +0800</lastBuildDate><atom:link href="https://lilfry09.github.io/tags/fullstack-agent/" rel="self" type="application/rss+xml"/><item><title>FullStack-Agent：全栈 Coding Agent 的训练与评测数据是如何构造的</title><link>https://lilfry09.github.io/ai/fullstack-agent-data-construction/</link><pubDate>Thu, 06 Aug 2026 20:00:00 +0800</pubDate><author>fry</author><guid>https://lilfry09.github.io/ai/fullstack-agent-data-construction/</guid><description><![CDATA[<p>很多“网页生成 Agent”看起来已经能做出漂亮页面：按钮能点，表单能提交，页面还能弹出“保存成功”。但如果继续追问两步，问题就会暴露出来：数据真的写进数据库了吗？刷新页面之后还能读回来吗？后端 API 是否真的存在？</p>
<p><code>FullStack-Agent</code> 这篇论文的核心贡献，正是把网站生成从“看起来像全栈”推进到“前端、后端和数据库都能被验证”。不过，论文里更值得仔细看的其实不只是 Agent 框架，而是它的数据构造方法：<strong>它如何把真实网站代码库转换成 Coding Agent 的训练轨迹，又如何构造能够识别假后端的评测数据？</strong></p>
<p>本文只讨论这两个问题。论文原文是 <a href="https://arxiv.org/abs/2602.03798" target="_blank" rel="noopener noreffer ">FullStack-Agent: Enhancing Agentic Full-Stack Web Coding via Development-Oriented Testing and Repository Back-Translation</a>，代码仓库见 <a href="https://github.com/mnluzimu/FullStack-Agent" target="_blank" rel="noopener noreffer ">GitHub</a>。</p>]]></description></item></channel></rss>