Feeding the output to another inference would probably reduce hallucination? but I have this impression that the models talking to each other don't need to know English, a byte stream would do. Just like the hidden layers of an ANN.