Named Entity Recognition (NER) plays a crucial role in structuring educational resources and enabling intelligent tutoring systems. However, the scarcity of annotated data and the prevalence of domain-specific entities pose significant challenges. This paper presents MAC-NER, a novel structured neural framework that combines multi-agent collaboration with external knowledge graphs to tackle these issues. Our approach decomposes the NER process into specialized agents for prompt generation, span extraction, and entity representation. To ensure global consistency, entity-type assignment is formulated as a linear optimization problem and solved via the Hungarian algorithm. Furthermore, a gated knowledge-enhancement module dynamically integrates entity-centric information from educational knowledge graphs. Extensive experiments on real-world datasets demonstrate that MAC-NER outperforms state-of-the-art baselines, exhibiting superior robustness and strong generalization in few-shot settings. The proposed system offers a practical and extensible framework for intelligent educational applications.