This is a problem we're looking at atm. We want to start running evaluations as we're having problems with consistency across releases even just for individual agents.