The design framework for pedagogical agents supporting constructivist learning with generative AI

Abstract

While generative AI lowers the barrier to creating pedagogical agents, unguided designs risk undermining learner agency and exposing learners to hallucinated content. This paper argues that a constructivist framing provides a strong theoretical basis for realigning pedagogical agent design to promote active knowledge construction and problem-solving. Guided by constructivist principles, we propose a design framework that (1) incorporates context-aware, in-the-moment guidance within authentic, open-ended learning activities; (2) supports the development of interpretation through productive cognitive conflict; (3) encourages collaboration and group discussions; (4) allows artifact co-creation with formative evaluation; and (5) fine-tunes cognitive apprenticeship strategies and intervention timing. We demonstrate how agents implement these principles by applying generative AI and multimodal learning analytics across various constructivist learning settings, including science labs, simulations, and computational modeling environments. We also highlight key safeguards essential for designing and deploying responsible pedagogical agents, including roles in collaboration, teacher oversight, prompt-before-answer rules, positive and open attitudes toward agent use, and maintaining activity authenticity. Grounding pedagogical agents in constructivist principles enables effective use of generative AI's capabilities while addressing its risks, thereby supporting deeper constructivist learning.