ETF Agent Skills Architecture Separates Methods from Execution
ETF Agent authors describe an architecture in which Skills define reusable research methods while models select and synthesize them, the Host controls capabilities and source authorization, deterministic programs handle calculations, and the Runtime manages budgets, execution, and recovery. The article also covers progressive loading, differing Skill visibility for Entry, Main, and Child agents, receipt-based verification, version governance, and a reproducible public v0.2 demo. The internal fund-research bundle and private-portfolio workflow remain vendor-described, while the public implementation supports only a narrower synthetic-guide example.
The article provides a concrete separation between reusable agent methods, permissions, deterministic computation, and runtime controls, while emphasizing that discovering a Skill does not prove it was read, executed, or adopted.
一个 Agent 已经能搜索网页、读取 PDF、调用 MCP、运行 Python，为什么还需要 Skill？
因为拥有工具，不等于掌握完成一类任务的方法。
假设用户问：“比较这两只指数基金的跟踪质量，再看看它们在我的组合里有没有重复暴露。”
搜索工具可以取得基金净值、基准与披露持仓；Python 可以计算主动收益波动和重叠比例。但工具不会替研究者决定：是否先核对份额、币种、分红口径和共同窗口？缺少私人组合授权时，是否应该停止计算个人重叠？两条序列无法对齐时，是否应承认指标不足？
在 ETF Agent 中，我们把这些可以复用的研究方法组织为 Skill，同时坚持一项分工：Skill 定义方法，模型选择并综合方法，Host 管实际能力和来源授权，确定性程序管计算，原 Runtime 管预算、执行和恢复。
这篇文章不仅讨论 Skill 文件结构，还沿一次研究解释：同一个 Skill 为什么在 Entry、Main 和 Child（Subagent）眼里不同；如何从目录变成实际读取、计算与交接；哪些行为可由回执确认，哪些不能靠目录或声明推断。