From Rule-Based to Goal-Oriented: A Systematic Review of Agentic AI in Education

The integration of artificial intelligence in education is rapidly shifting from reactive language models to Agentic AI—autonomous, multi-agent ecosystems capable of sustained reasoning, dynamic tool use, and long-horizon goal planning. While these systems offer new opportunities for personalized tutoring and dynamic assessment, their unchecked deployment risks prioritizing computational efficiency over genuine cognitive development. This systematic review synthesizes 53 studies (2022–2026) following PRISMA guidelines to map the current landscape of agentic educational technologies. We analyze the literature across five core domains: specialized multi-agent architectures, pedagogical applications, dynamic knowledge tracking, theoretical alignment, and new evaluation methodologies. Specifically, we scrutinize the emerging theory of Agentivism and address the long-term risks of automated cognitive offloading. A critical evaluation of prominent multi-agent systems reveals significant methodological gaps in the current state of the art, notably a widespread over-reliance on small synthetic benchmarks, unfair baseline comparisons, and a severe lack of longitudinal testing in authentic classrooms. In addition, we examine emerging socio-technical vulnerabilities, such as epistemic lock-in and the need for human-in-the-loop oversight. To address these challenges, we propose a research roadmap emphasizing the rigorous empirical validation of cognitive impacts, the development of standardized educational benchmarks, and the implementation of neuro-symbolic governance to ensure system safety. Ultimately, this review establishes a guide for the responsible and pedagogically sound development of autonomous learning companions.