Neural machine translation has significantly improved translation accuracy, attracting various sectors of society. A highly accurate neural machine translation requires a large amount of high-quality bilingual corpus. Sharing bilingual data among different organizations is challenging due to copyright and confidentiality concerns. Federated learning resolves this by allowing multiple clients to collaboratively improve a translation model while maintaining the privacy of their individual datasets. However, the varied nature of the data each participant holds can diminish the model’s accuracy, and integrating all translation models does not ensure improved accuracy for all. To solve this problem, we propose a multi-agent system that uses deep reinforcement learning to integrate translation models during the federated learning process. The system dynamically selects partners for each aggregation phase, with agents developing strategies through deep reinforcement learning to identify the most suitable partners for collaboration. We have validated our method using Wikipedia data from three distinct domains, and it has demonstrated superiority over traditional methods. Additionally, we have confirmed that our method increases the translation accuracy for each client by 33.9% compared to the current federated learning approach.