Approaches and emerging trends in multi-agent autonomous AI systems for education innovation in Vietnam

Agentic artificial intelligence (AI) systems that act as autonomous agents are rapidly evolving due to the explosion of large language models (LLMs) [1]. This paper presents a comprehensive analysis of major approaches in the field of Agentic AI, including the large visual language model (LVLM), the React and Plan-and-Execute agent architectures, the smolagents library with a “coding-first” orientation, tool invocation techniques that extend the capabilities of LLMs, the visual Agentic AI model with multi-agent coordination capabilities, as well as scientific agent systems such as AI Scientist and the AgentRxiv collaboration platform. We analyze the characteristics of each approach, including representation models, advantages, limitations, and integration capabilities, for building intelligent agent systems that aim for AGI. The paper proposes an integrated scheme that leverages achievements from multimodal capabilities, multistep reasoning and planning, multi-agent coordination, and research automation, laying the foundation for a new generation of autonomous AI agents. Finally, we discuss the potential applications of Agentic AI in the context of Vietnam, especially in education, scientific research, and technology development, and provide recommendations for domestic developers and researchers.