The concept of group work has been at the center of higher-education pedagogy since time immemorial, but AI-powered platforms have only started to pursue other functions besides grading and content delivery. This narrative review explores agentic AI, i.e. systems that make semi-autonomous decisions, as virtual partners, as transforming collaborative learning between 2018 and 2025. Using the social learning theory, cultural intelligence frameworks, and cybernetic feedback models, we discuss how agentic AI can intervene in group activities as a co-learner, facilitator, and mediator. Empirical research indicates that such systems have the potential to expand engagement, redistribute the speaking turn and even enhance more profound thought, yet they also bring about the issues of trust, bias, and digital inequity. In case examples, such as voice-enabled AI speakers, adaptive LMS platforms, and conversational chatbots, we can find an example of the promise and the real-life constraints of the implementation. Lastly, we address what these changes in the role of instructors, institutional policy in the area of ethics and access, and major directions of future research, including emotionally intelligent AI agents and multi-agent ecosystems. This review provides an insight into how educational technologies could be structured to enhance, instead of emulating, the strength of human collaboration by concentrating on the human and social aspects of AI integration.