People already do this with multi agent workflows. I kind of do this with local models, I get a smaller model to do the hard work for speed and use a bigger model to check its work and improve it.