Large Language Model (LLM)-based AI Agents, such as ChatGPT, Claude, and Gemini, have demonstrated strong potential in natural language processing and intelligent tutoring. However, vocabulary acquisition among primary school students (ages 6–11) remains a persistent challenge due to limited motivation, repetitive drills, and a lack of contextualization. To address this gap, we propose an educational game designed for children that integrates an AI Agent as an intelligent orchestrator of the learning process. The agent leverages contextual understanding, memory, and tool integration to support error-tolerant multimodal interaction (text and speech). By analyzing data collected during gameplay, the system dynamically tracks learning progress, adjusts difficulty, modifies game flow, and expands vocabulary content without manual reconfiguration. Experimental evaluations demonstrate that the proposed architecture achieves a 123.1× speedup in response latency compared to pure-LLM baselines, while the adaptive mechanism yields a 52% increase in vocabulary acquisition and a 90% user retention rate. This approach overcomes the limitations of traditional vocabulary learning systems, which often rely on static databases and rigid structures. The anticipated contributions of this work include: (i) enhancing vocabulary retention through gamification, (ii) increasing engagement and motivation among primary school learners, and (iii) demonstrating the feasibility of AI Agents as adaptive and personalized coordinators in educational environments. Ultimately, this work introduces a novel agent-based learning paradigm that aligns with the language learning needs of children in the AI era, paving the way for more interactive, personalized, and sustainable vocabulary learning systems.