Large language models have shown potential in educational content generation, but face challenges in ensuring the appropriateness. This research proposes a “generation - judge” multi-agent system to produce multi-modal educational content tailored to preschool children. Specifically, grounded in Vygotsky’s Zone of Proximal Development, the generation agent produces multi-modal content aligned with the Early Years Foundation Stage (EYFS), and the judge agent evaluates the content’s appropriateness based on the same EYFS framework. In the child assessment stage, the generation agent generates evaluation content according to the American Academy of Pediatrics (AAP) clinical practice guidelines, while the judge agent assesses the generated content against the AAP guidelines. Through controlled experiments with intervention and control groups, we assess the impacts of the generated content on young children’s holistic development. The findings aim to provide empirical support for the use of LLMs in educational content generation and offer practical solutions to enhance learning outcomes.