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?