ARGUS: An MCP-Based Agentic LLM Architecture for Distribution Grid Decision Support

Abstract

Modern power distribution grids are becoming increasingly complex due to dynamic network configurations, evolving customer behaviors, and fluctuating power demands. These factors introduce indirect spatial and temporal dependencies that critically affect voltage profiles, line loading, power losses, and overall system reliability. Consequently, grid operators require interactive decision-support tools to evaluate grid behavior under diverse operational scenarios. This paper presents ARGUS, an interactive agentic system for power distribution grid analysis based on the Model Context Protocol (MCP). The proposed framework enables operators to explore how modifications in grid configuration and load conditions influence key operational indicators, including bus voltages, line thermal limits, and network losses. By isolating the Large Language Model (LLM) to orchestration, intent recognition, and knowledge retrieval, all physical calculations are grounded in validated AC power-flow solvers and surrogate models. Demonstrated on the IEEE 33-bus system driven by empirical demand and DER profiles, ARGUS accelerates the detection and interpretation of operational bottlenecks, providing an auditable and trustworthy decision-support environment for modern distribution networks.