Honestly, I think complex architectures are best demonstrated as diagrams—and those can be developed in an agile fashion. Stable, well-thought-out architectures can’t be slapped together without nice diagrams. There’s a ton of folks who just “start coding” to get a feature going, but when someone else takes over the project, how are they to learn the code? Diagrams are always the best way for me—and there’s limits on what doxygen says, depending on how bad the implementation is.

Main point of UML is to tackle both diagramming/architecture AND forcing basic coding to reflect the diagrams. It forces code and documentation to both reflect the architectural truth.

This doesn’t have anything to do with agile methodologies, as any task can follow agile workflow.