Supervision is important to vertical dilemmas. Interaction rules of supervision and violation were constructed by multiagent modeling and simulation. The evolving mechanism was designed using genetic algorithm and the evolving process was simulated on the swarm platform. It is concluded that under “limited rationality” and “incomplete information”, the supervisor must obtain almost the “complete information” of the behavior of participants, apart from making their expected return of obedience greater than the expected return of violation, to eliminate the behavior of violation. The progress of the evolution can be accelerated by harsh punishments