Automated Planning in Intelligent Distributed Systems Using a Multi-Agent Approach Based on LLM

The paper presents a novel approach to automating management of intelligent distributed information systems (DIS) using a hybrid multi-agent architecture with AI agents based on large language models (so-called LLM agents). The proposed architecture consists of a two-level planning system, which includes a manager agent for centralized global task decomposition and executor agents for decentralized local task assignment and computation optimization across individual clusters of the intelligent distributed system. To enhance agent interaction efficiency, the study proposes innovative methods for model access to long-term and short-term memory and tool utilization through code generation and execution, using the specialized prompt engineering method CodeAct. The CodeAct's high performance allows to create new agents, interacting with environments, execute interpreted code, and are able to collaborate with users using natural language. The proposed approach potentially provides architectural agnosticism regarding distributed system structure, allowing work with complex heterogeneous DIS and task types through specialized management components, memory access, and programmatic instruction execution generated by agents. The work includes descriptions of necessary skills and properties of the base large language models, the structure of two agent types, and the set of tools required for effective interaction of intelligent system’s agents with each other and the external environment, including the managed DIS. The paper presents architectural diagrams of the manager agent and executor agents, describing the internal structure of all components and the mechanisms by which they interact. The research explains the promising potential of large language models and advanced AI agents for intelligent planning in distributed computing environments, revealing the potential of artificial intelligence in automating complex management processes.