This paper proposes an innovative approach to training Minecraft Battle Bots capable of competing against human players. We designed a lesson-based training framework (Lesson 1 to Lesson 6) that progressively increases task complexity, enabling agents to learn foundational skills and gradually develop the ability to handle complex combat scenarios. Experimental results demonstrate that these lesson designs are critical to the model’s training success, as agents fail to achieve learning objectives without them. Furthermore, we introduce an Online Training Framework that enables real-time online learning, allowing agents to continuously optimize their behavior and adapt to diverse player strategies through real-time interactions. Experimental evaluations confirm this framework’s real-time capability and stability, ensuring minimal inference delay during training and enhancing the fluidity of interactions between players and agents. In summary, this paper integrates a lesson-based training framework with an online real-time training system, significantly improving the efficiency and adaptability of Minecraft agents.