提问 HN：基于 Agent 的质量保证（QA）可行吗？
1 分•作者: straydusk•2 个月前
由于诸多原因，测试驱动开发（TDD）已成为人工智能编码代理的默认工作流程。<p>我最近开始做的一件事，我非常喜欢，就是在构建一个功能时，我先写出验收标准，然后确保代理拥有完成质量保证（QA）并自行验证所需的一切。与之前的最大区别在于，我不再自己进行质量保证（虽然我仍然会做，但几乎总是有效的），而是让代理自己执行整个流程并验证验收标准。这需要更多的工作，因为我经常需要设置多方通信协议（MCP）、账户、凭证等——但输出结果更好。<p>我的问题是……大家对此有什么好的方法吗？这是一种常见的工作流程吗？我看到有人将这种讨论称为“工具工程”；也有人称之为“反馈循环工程”；我看到一些初创公司（Shiplight AI、Autosana、Ranger）也在做类似的事情。我正在努力寻找更多关于这方面的讨论和最佳实践，并了解其他人是如何思考这个问题的。
查看原文
TDD has become the default workflow for AI coding agents, for a lot of reasons.<p>Something I recently started doing that I really like is, when building a feature, I write acceptance criteria and then I make sure the agents have everything they need to do QA and verify it themselves. The main difference from before is that instead of QAing it myself (which I still do but it's almost always working), the agent can exercise the whole flow itself and verify the acceptance criteria. It takes more work because I often have to set up MCPs, accounts, credentials, etc - but the output is better.<p>My question... do people have good approaches to this? Is this a common workflow? I see some of this conversation called harness engineering; I see some called feedback-loop engineering; I see some startups (Shiplight AI, Autosana, Ranger). I'm trying to find more discussion & best practices here, and see how others are thinking about this.