Abstract
Mathematical ability is widely recognized as a core indicator of students' cognitive development and academic achievement. However, in classroom environments where artificial intelligence (AI)-supported instruction is increasingly integrated into routine teaching practices, how mathematics anxiety shapes the development of mathematical ability over time remains insufficiently understood. To address this gap, this study examines the longitudinal relationship between state mathematics anxiety and mathematical ability among primary school students in AI agent-supported instructional contexts, with particular attention to the moderating role of executive function. Using a convenience sampling approach, a total of 494 students from Grades 1, 3, and 5 participated in routine AI agent-supported mathematics instruction over a six-month period, with AI-supported lessons implemented approximately twice per week on average. A two-wave longitudinal design was adopted, and cross-lagged panel analysis was conducted to examine the directional relationships between the key variables. The results indicated an asymmetric relationship, in which mathematics anxiety significantly predicted subsequent mathematical ability, while the reverse effect was not supported. Specifically, higher levels of mathematics anxiety predicted subsequent declines in overall mathematical ability, mathematical computation ability, and logical-spatial ability, whereas prior mathematical ability and its subdimensions did not significantly predict later mathematics anxiety. In addition, executive function moderated the relationship between mathematics anxiety and mathematical ability outcomes in a domain-specific manner. Higher levels of executive function strengthened the negative association between mathematics anxiety and both overall mathematical ability and mathematical computation ability, but not logical-spatial ability. A similar pattern was observed for inhibitory control, whereas working memory and cognitive flexibility did not show significant moderating effects. This study contributes new longitudinal evidence on students' mathematical development in AI agent-supported classrooms and advances current understanding of how mathematics anxiety and executive function jointly shape learning outcomes, while also informing the design of more responsive instructional support.