The later takes too much mental ressources, the same when reviewing truly the code generated by the former.

I generally started by reviewing but after a while (maximum in hours), I just can't keep up and resort to LLMs as sole reviewers.