I wonder if using thinking models would work better here. They generally have less variance and consider more options, which could achieve the same goal.