Visual approaches are fundamentally limited by the expressiveness of visual constructs, and the available screen real-estate
3D modeling programs are a visual approach to 3D modeling without having to type the code for a box to render it on screen yourself. I don't have to learn DirectX or OpenGL to model a cube. I can just hit the "Cube" button.
This example makes 3D modeling sound trivial, but we all know that 3D modeling programs are very powerful. Why should everyone have to learn OpenGL and such to create a 3d model? Even though these 3d modeling programs are created textually, we interact with them visually. It would be the same for a node based system in programming. There wouldn't be any additional limits by simply changing the way we interact with code (which is really what is happening).
So I'd have to disagree that Visual approaches are fundamentally limited in the way you describe.