2 篇博文 含有标签「agent-skills」

最近，我们一直在尝试优化 Better Harness 的 SKILL 自动沉淀能力：从 Agent 的真实会话中识别重复出现的工作路径，再判断其中哪些经验值得进一步沉淀成可复用的 SKILL。真正做起来以后，我们发现这件事远比“把一段 Session 分析一遍”复杂得多。

对于一次软件开发任务来说，Agent 的行为并不是孤立发生的。它从一个需求或者用户故事开始，经过对需求的理解、上下文探索、代码修改和验证，最终才形成一次可以被评审的代码贡献。只看中间的 Session，我们能看到 Agent 做了什么，却很难判断这些行为为什么发生，又有哪些行为真正进入了最后的交付。

因此，我们开始把一次 Agent 的交付理解成一条连续的链路。现在，只需要在项目目录执行：

npx @qoder-ai/better-harness inspector

就可以生成一个本地、只读的 Harness Inspector 页面，把当前项目中的 Agent Session、文件活动和 Git Commit 放到同一个交互界面里。

Writing a capability into a SKILL.md and packaging it as an Agent plugin is not hard. The hard part comes later: once that capability is invoked over and over by different users, in different projects, and on different Agent hosts, how do you guarantee that it is still triggered, executed, and verified correctly? And how do you prove that a change made it better rather than worse?

Drawing on how Better Harness is actually developed, this post walks through five engineering practices - spec-driven behavior, context orchestration, deterministic verification, behavioral evaluation, and the evidence loop - that move a plugin capability from "it works when I run it" to a software asset that is verifiable, maintainable, and safe to evolve.