Automating Abaqus scripting for cracked reinforced-concrete (RC) analysis remains challenging due to solver-specific APIs, topology-sensitive geometry/partition operations, and fragile post-processing that can yield either runtime failures or deceptively “successful” runs that do not satisfy engineering requirements. This paper presents AutoFEA, a solver-in-the-loop multi-agent framework that converts natural-language structural analysis specifications into executable Abaqus/CAE noGUI Python workflows and enforces requirement-complete convergence. AutoFEA structures each task into an explicit, checkable specification via Retrieval-Augmented Reasoning Enhancement (RARE), then synthesizes multiple candidate modelling routes with multi-judge pre-screening before solver execution. During execution, failures are classified into a descriptive error taxonomy and routed to category-specific repair agents with conditional retrieval gating over solver documentation and verified code patterns. After solver-level success, an output-aware Auditor verifies deliverable integrity and requirement coverage, bridging the gap between runnability and engineering adequacy. To evaluate task-level adequacy, CrackedRC-Suite is introduced as a 24-task benchmark spanning three difficulty levels, covering intact beams, seam/CZM/XFEM crack modelling, and advanced contact and multi-step nonlinear settings. Experiments show that both Zero-Shot generation and standard RAG fail to produce executable scripts (SSR = 0%). An ablated variant achieves SSR = 100% but leaves a large adequacy gap (overall TCR = 62.5%). With auditing enabled, AutoFEA reaches SSR = 100% and TCR = 100% across all levels, demonstrating reliable natural-language-to-verified FEA automation for construction-oriented RC modelling.