Архітектура Корпоративної Екосистеми Штучного Інтелекту У Цифровій Трансформації Університету

Rapid digitalization in higher education, combined with the need for universities to operate under geographically distributed and relocated conditions, highlights a crucial shift from rigid hierarchical management to flexible, decentralized systems. This paper substantiates the architecture of decentralized university management models driven by multi-agent systems and large language models (LLMs), presenting its architectural justification and functional implementation within a higher education institution. The study details theoretical foundations of utilizing multi-agent systems, orchestration protocols, Retrieval-Augmented Generation (RAG), and vector databases to optimize university workflows under conditions of limited observability and distributed IT infrastructure. The methodology rests on a two-tier multi-agent architecture, where dynamic request routing is achieved through a hierarchical orchestration model using AgentSpec formal specifications, while action consistency is maintained by a central algorithmic coordinator. Findings indicate that deploying a standalone autonomous agent for each of the university’s 24 departments is inefficient due to risks of algorithmic conflicts, redundant queries to the vector knowledge base, increased LLM hallucination rates, and context window constraints. To address these limitations, a coordinated multi-agent architecture was implemented at Kherson State University. This framework consolidates operations into five core macro-agents—academic affairs, administrative management, knowledge management, educational quality assurance, and communications—managed by a Central Orchestrator routing queries, controlling access privileges, and synthesizing analytical outputs. The paper identifies key organizational and technological hurdles, focusing on maintaining high levels of trust, decision verification, and role-based access control during delegation. Finally, it outlines the roadmap for evolving this framework into a university digital twin capable of predictive modeling and automated process support. These findings offer practical value for higher education institutions deploying digital ecosystems under human resource, infrastructural, and security constraints.