In multi-agent reinforcement learning (MARL), coordinating collaboration during adversarial tasks remains challenging. Existing methods like opponent intention modeling and centralized training with decentralized execution (CTDE) either require global information or struggle with static task flows, limiting real-world applicability. To address this, we propose the Multi-Agent Dynamic Intention Recognition (MADIR) algorithm. MADIR uses an action prediction module to infer allied agents’ behaviors from local observations and introduces a dynamic learning rate adjustment to enhance convergence stability. Experiments on the StarCraft II challenge and a custom Unity environment show MADIR outperforms state-of-the-art CTDE methods in task success rates.