Agentic AI as Process Scaffolding in Undergraduate ESL Instruction: Effects on Proficiency, Engagement, and Self-Regulated Language Learning

Abstract

Recent AI-assisted language learning research has focused mainly on reactive chatbots that respond to prompts but offer limited support for planning, monitoring, and follow-up. This study examined whether a bounded agentic AI environment designed as process scaffolding would be associated with stronger ESL outcomes than a reactive AI interface. Using an explanatory sequential mixed-methods, quasi-experimental design, 128 undergraduates from four intact academic communication sections at two private universities in the Philippines participated in a 12-week intervention. Students in the experimental condition used an Agentic English Coach that supported weekly goal setting, task decomposition, reminders, reflection prompts, adaptive resources, and teacher escalation; the comparison group used a reactive chat interface to complete the same course tasks. ANCOVA results controlling for pretest scores showed that the agentic AI group outperformed the reactive AI group in posttest English proficiency, learner engagement, self-regulated language learning, and agentic engagement. Within the experimental group, plan completion, reflection completion, revision cycles, and speaking practice turns positively predicted posttest proficiency. Interview data indicated that agentic AI reduced task uncertainty, supported sustained practice, made progress more visible, and shifted teachers toward orchestration and oversight of feedback rather than constant procedural prompting. The findings suggest that AI is most pedagogically valuable in ESL instruction when it structures learning processes over time rather than only generating on-demand responses.