As AI technology advances rapidly, the use of generative AI in education is expanding. While generative AI serves as a powerful tool for extending learners' capabilities, it also poses the risk of undermining learners' sense of responsibility for lear...
As AI technology advances rapidly, the use of generative AI in education is expanding. While generative AI serves as a powerful tool for extending learners' capabilities, it also poses the risk of undermining learners' sense of responsibility for learning through excessive dependence on AI. In this context, the importance of student agency—learners' capacity to autonomously regulate their learning processes and make decisions while receiving AI support—has become increasingly prominent. However, existing research has been limited to exploring the effects of AI on learning or discussing the design of learning environments to support student agency only at a theoretical level.
Therefore, this study aimed to design and implement an LLM-based AI agent system that supports student agency in generative AI-integrated learning environments and to explore its influence on high school students during data analysis projects. The research questions were: (1) What are the structures and functions of AI agents that support agency in AI-utilizing data analysis? (2) How do learners respond to these AI agents? (3) Are the structures and functions of these agents effective?
Using a Design-Based Research (DBR) methodology with two iterative cycles, 15 high school students from a statistics club participated. In Cycle 1, a system comprising a "Data Analysis Agent" and a "Learning Agent" was developed using LangChain. Results from usability surveys and FGIs showed high motivation and self-efficacy but low levels of planning and prompt literacy, with students noting a need for clearer structures. Reflecting this, Cycle 2 introduced a structured three-phase workflow (Planning-Execution-Reflection) and redesigned the system using LangGraph to integrate state management across stages. Data were collected via pre-post tests, in-depth interviews, and interaction logs.
The results of paired-samples t-tests revealed significant improvement in the agentic action dimension (t = –2.858, p = .013). The competence dimension (t = –2.120, p = .052) showed a trend approaching significance, while the psychological dimension (t = –0.947, p = .360) showed no significant change. In-depth one-on-one interviews revealed that the AI agents supported students' agency and strengthened their sense of learning responsibility. Furthermore, this experience transferred to general-purpose AI use, enabling students to utilize generative AI like ChatGPT more effectively in other contexts.
Based on these findings, three design implications were derived. First, AI agents should act as active orchestrators that track progress, where explicit planning phases are crucial for goal-directed action. Second, agent design should prioritize affordances that promote agency; specifically, the "unfriendliness" (intentional restraint) of the data analysis agent enhanced students' ownership of their learning. Third, learners' beliefs about AI significantly impact agentic AI use; students perceiving AI as a collaborative partner reported competence gains, while those seeing it as a mere tool expressed skepticism.
This study bridged the gap between theory and practice by establishing the support of student agency as an explicit goal and designing and implementing a system utilizing state-of-the-art LLM-based agent architecture. The significance of this research lies in proposing a new direction for human-AI interaction in generative AI-integrated learning environments.