Open-source large language models (LLMs) are increasingly explored in educational contexts due to their transparency, adaptability, and alignment with institutional governance and equity considerations. Despite growing interest, empirical research on how open-source LLMs are deployed in education and what evidence currently supports their integration remains limited and fragmented. This paper presents a state-of-the-art narrative review of peer-reviewed, human empirical studies examining the use of open-source LLMs in education. Guided by three questions, the review synthesizes how open-source LLMs are deployed across instructional contexts, what learner-related evidence is reported, and how teachers engage in human–AI collaboration. The reviewed literature is concentrated in higher education, particularly within computer science and programming domains, with applications focused on post-class tutoring, guidance, and formative feedback. Learner perceptions are generally positive, but evidence linking open-source LLM use to measurable learning outcomes remains emerging and inconsistent. Through interpretive synthesis, the review articulates a four-role model—Designer, Facilitator, Monitor, and Evaluator—that captures how teacher agency is enacted across AI-supported instructional workflows. This review maps recurring orchestration dimensions, decision points, and tensions that characterize early implementations, and it proposes a minimal orchestration reporting scaffold (configuration, boundaries, logging, adjudication) intended to support auditability and cross-study comparison as the empirical base develops.

This study explores the integration of a multi-agent AI system designed to support self-regulated learning (SRL) in computer science education. Building on preliminary work that introduced a Teaching Assistant AI (TA-AI) for real-time scaffolding, an Analytics-AI for pattern recognition, and an AI-Improver for iterative refinement, this study evaluates the system’s empirical impact on student help-seeking behavior. Guided by fundamental SRL principles, this study analyzed data from 38 graduate students across 109 help-seeking attempts to investigate the factors influencing prompting strategies during system implementation. Findings showed that students utilized the TA-AI as a strategic safety net. Questions where students sought help were associated with significantly lower correctness (OR = 0.35, p < .001). A “sweet spot” in quiz difficulty was identified where the TA-AI is most effective in activating metacognitive prompting: medium-difficulty questions significantly increased the likelihood of