Many people ignore the code execution environment that Skill relies on.
This is not just a purely theoretical architecture concept, but also provides specific solutions at a truly operational level.
When this solution is actually implemented, the upper-layer manager agent and subAgent can be implemented and defined purely with logic.
However, when it comes to the tool layer, it will involve a lot of calls, data transfer, tool adaptation, environment processing, and runtime issues.
Therefore, if you really want to implement and maintain a pure AI Native architecture, add a layer of SKills as a bridge between SubAgent and Tool.
Do not let SubAgent operate Tools.
SubAgent only focuses on Agent logic.
This is also a common decoupling design.