That is what’s already been done under the term "multi-agent". This paper argues that there’s no need for any such message-passing or context sharing, you just literally run the same query several times on the same model, fully independently, and then pick a "typical" reply according to some similarity metric.