Recent advances in educational technology highlight an urgent need to understand how emerging Artificial Intelligence (AI) agents can transform learning environments. Traditionally, agents are defined as artificial entities capable of sensing their environment and encompassing functions like perception, reasoning, decision making and action. Historically grounded in symbolic AI and rule-based reasoning, these systems often lacked the flexibility to process complex semantic contexts.

The advent and rapid development of Large Language Models (LLMs), particularly since 2022, have driven a paradigm shift from pre-programmed models towards LLM-based agents, revolutionising the capabilities of traditional agents, especially in terms of complex context comprehension and human-like language generation. LLMs enable AI agents to act as “decision centres”, equipping them with advanced linguistic understanding, generation and reasoning abilities. This enables more sophisticated, human-like interactions, including engaging in in-depth dialogues, answering questions, clarifying concepts, participating in debates and explanations, and efficiently collaborating with humans to develop solutions. Moreover, LLM-based agents can handle complex natural language inputs and dynamically adapt their responses based on context and user input, offering more personalised and intelligent services. In addition to the evolution of the technologies used for agents, their design patterns are also evolving. Several studies suggest that to operate effectively in complex environments, agents should be autonomous, proactive and social.

While individual LLM-based agents, such as well-known conversational chatbots (e.g., ChatGPT, Claude), are already widely used to respond to requests and generate human-centred content and feedback, recent trends highlight their integration into Multi-Agent Systems (MAS). Functioning as communities, these systems enable collaboration among multiple agents to address complex tasks that exceed the capabilities of any single one.

In the field of education, recent literature suggests that agents play a crucial role in various areas, including personalised learning support, intelligent tutoring and Q&A, virtual labs and simulation environments, personalised content generation, learning process monitoring and analysis, as well as educational games and game-based learning.

This work aims to investigate how the evolution of LLM agents is transforming educational practices, outlining current trends in LLM-MAS within educational contexts and seeking to understand the theoretical foundations underpinning these applications. Specifically, the analysis draws upon Computer-Supported Collaborative Learning (CSCL) and Social Learning Theory perspectives. By synthesising these theoretical foundations with technological innovation, the aim is to provide a deeper conceptual understanding and practical guidelines for integrating AI agents into educational contexts.

Although such integration offers significant advantages in personalisation and efficiency, it is crucial to balance technological support with fostering students’ critical thinking and problem-solving abilities, mitigating risks of over-dependence. Overall, understanding the integration of LLM-based agents within collaborative educational settings holds the potential to fundamentally reshape instructional practices, paving the way for more inclusive, adaptive and engaging learning environments.

Keywords: Educational Technology, AI Agents, LLM, Multi-Agent Systems, CSCL.