Abstract
Artificial intelligence (AI) is rapidly transforming education, shifting it from static courses to personalized, adaptive learning ecosystems. According to international organizations (including UNESCO), AI is becoming an infrastructural element of higher education; within this frame, intelligent agents (IAs) serve as a mechanism for integrating pedagogical objectives, learner data, and real-time adaptation strategies. This article aims to systematize the architectural and functional principles of using intelligent agents in educational technologies and to analyze implementation practices within contemporary AIEd. We trace the evolution from monolithic Intelligent Tutoring Systems (ITS) to distributed Multi-Agent Systems (MAS) and dialog agents powered by Large Language Models (LLMs). Empirical findings on the effectiveness of classical ITS are synthesized and compared with emerging practices of LLM-based agents on mass-scale platforms. The study’s novelty lies in an analytical comparison across three levels—architectural (ITS/MAS), instrumental (dialogic and analytic functions), and institutional (policies and deployment metrics)—grounded in evidence from 2023–2025. In addition, we formulate methodological guidelines for responsible adoption (explainability, fairness, and data protection) to balance automation with pedagogical oversight and to define requirements for scalable, ethical, and transparent learning ecosystems.