Hate speech detection is a key element of maintaining a healthy discourse environment for social media platforms, particularly for low-resource languages like Vietnamese. In our paper, we propose a new approach drawing inspiration from multi-agent debate to improve the accuracy of discrimination between toxic and non-toxic comments. The system consists of two independent large language model agents that initially produce independent decisions, then undergo a formal debate arguing their sides, and a neutral judge agent finally produces the decision. We try the system out on the dataset of the ViCTSD and demonstrate that our approach attains 89.80% accuracy and 79.66% macro F1-score, outperforming the standard single-model baselines. Our experiments demonstrate that the incorporation of the debate scheme helps the agents better put the content in context and reduces the misclassification arising from ambiguous phrasing or latent aggressivity. Our paper demonstrates a potential direction towards leveraging cooperative AI agents to enhance the reliability and interpretability of systems for the moderation of toxic content.