Our agent re-checks its own answer until the critic pas...

A strong answer recognizes that rising confidence is not evidence of improved accuracy. Because the model is conditioning on its own previous text, it is not incorporating new information. You must measure accuracy as a function of the round index on a labeled set to determine if the loop actually provides any performance gain before deciding to ship.