Effect of Prepared Skills and Tools on AI Agent Performance in Power System Steady-State Analysis

The article proposes an AI agent performance architecture based on a large language model (LLM) and intended for steady-state analysis of electric power systems in the RastrWin3 software and computing system. Engineering skills define the sequence of actions and interpretation rules; software tools perform standard operations on the calculation model, and the Model Context Protocol (MCP) provides a standard interface through which the agent invokes these tools. The RastrWin3 system produces numerical results. Two isolated runs of GPT-5.6 Sol in Codex, with reasoning effort set to high, are compared: an AI agent without prepared skills and tools, and an AI agent with prepared skills and tools. Both agents completed the full N−1 study of 787 standard contingencies on a modified IEEE 300-bus test system. The agent without prepared skills and tools misinterpreted the RastrWin3 calculation fields and independently selected voltage limits that were inconsistent with the adopted engineering practices. The agent with prepared skills and tools correctly applied the current and voltage assessment criteria; its results were confirmed by a repeated RastrWin3 calculation. Thus, the prepared skills and tools improved the interpretation of the complete N−1 study results.