CONTROLLED APPLICATION OF MULTI-AGENT SYSTEMS FOR TASK PLANNING IN INTELLIGENT DISTRIBUTED INFORMATION SYSTEMS

Abstract

The growing need for resource planning automation in distributed information systems creates new challenges regarding the balance between the efficiency of autonomous solutions and the need to ensure expert control. Traditional approaches to automatic planning, including static algorithms and classic machine learning methods, do not fully utilize the potential of modern artificial intelligence technologies, while fully autonomous systems based on LLM agents pose high risks when working with real computing resources. This paper proposes a hybrid multi-agent architecture for automatic task planning in distributed information systems with an integrated human-in-the-loop component. Architecture combines the logical reasoning capabilities of large language models with the possibility of human control, providing centralized planning through a manager agent and decentralized execution through executor agents. A key feature of the proposed approach is a multi-level validation system for the generated code, which includes static analysis, the possibility of pre-execution, and mandatory confirmation by a specialist at critical decision-making stages. The research methodology is based on the principles of hybrid planning, which allows for the effective use of the advantages of LLM agents while minimizing the risks of their application in real computing environments. The system’s application is clearly limited to non-critical intelligent data processing tasks, including machine learning, data analysis, natural language processing, and computer vision. The proposed architecture solves the problem of balancing the need for resource planning automation with the need to ensure safety when working with real distributed infrastructures, creating a theoretical basis for the controlled implementation of agent technologies in the practice of managing distributed systems.