Agentic AI as a dual-role lecturer and teaching assistant: effects on learner autonomy, self-regulated learning, and intrinsic motivation in university-level EFL education.

Large language model (LLM)-powered agentic systems are increasingly deployed in higher education, yet their effects on learner autonomy and self-regulated learning (SRL) remain empirically underexamined. This 16-week quasi-experimental study evaluated a dual-role Agentic AI-an AI Lecturer paired with an adaptive AI Teaching Assistant-embedded in Uedu.tw, a university EFL platform in Taiwan, drawing on Self-Determination Theory (SDT) and Vygotsky's Zone of Proximal Development. Participants were 120 Taiwanese B1 undergraduates (60 experimental, 60 control) assigned by intact class with a single instructor and identical curriculum. ANCOVA showed greater experimental-group gains in learner autonomy (partial η2 = 0.247), SRL strategies (partial η2 = 0.215), and intrinsic L2 motivation (partial η2 = 0.183), with no parallel effect on extrinsic goal orientation-a dissociation consistent with SDT. Because the experimental condition combined AI access with a Socratic-dialogue task format, these effects index the AI-supported condition as a whole and cannot be attributed to the AI system in isolation. Learner-level analysis of 6,235 AI Teaching Assistant queries showed that 54 of 60 experimental learners (90.0%) shifted monotonically from answer-seeking toward metacognitive and elaboration-seeking queries, which we treat descriptively rather than as evidence of internalized self-regulation. Reflexive thematic analysis of 16 interviews yielded three illustrative themes. The findings inform agentic-AI design in Confucian-heritage EFL contexts and underscore the need to distinguish effects with agentic technology from effects of it; whether within-intervention gains persist after scaffolding is withdrawn remains open.