Enhancing reliability and automation of LLM-based structural analysis using a hybrid multi-agent pipeline

Large language models (LLMs) can assist engineering workflows, but their direct application in structural analysis is fundamentally limited by numerical inconsistencies, prompt sensitivity, and context degradation. To address these bottlenecks, a verification-and-refinement driven multi-agent framework is proposed. The five-stage pipeline uses an explicit verify–correct loop to improve engineering consistency and incorporates a Model Context Protocol (MCP) hybrid path. This MCP architecture routes highly complex numerical structural calculations to a deterministic external solver under predefined trigger conditions. The methodology was evaluated on two cases: (A) a benchmark frame collapse-analysis problem resolved purely within the LLM verify–correct loop, and (B) an eight-story steel moment-resisting frame where trigger-based routing invoked MATLAB for high-DOF second-order analysis. Across the evaluations, the iterative pipeline increased verification pass rates, with the largest gain occurring during the first verify–correct iteration. Staged prompts and structured JSON/Markdown handoffs also limited context inflation. In Case B, the MCP-delegated numerical solve returned reports that satisfied drift and equilibrium requirements. These results suggest that an iterative verification pipeline combined with policy-guided external-solver delegation can improve reliability in complex structural engineering problems.