Dataflow models are definitely convenient. One of the nice thing about QuartzComposer is the ability to directly "probe" connections between patches to see what data was flying across them, whether this was text or something richer (e.g. graphics).

On the other hand, dataflow models are definitely limited in their abstraction capabilities, which is in opposition of directness. They work well in shells but have a harder time being used in large scale projects. I've never been able to make them scale past a certain point.