Reasonably reasoning AI agents can avoid game-theoretic failures in zero-shot, provably