Presumably all these agents would have been trained on different data, with different viewpoints? Otherwise, what makes them different enough from each other that such a "conversation" would matter?

Different skills or plugins, different views and different tools for the analysis of the same object. Then the debate starts.