Multi-agent systems improve structural quality of AI-generated sustainable product design specifications

Abstract Large Language Models (LLMs) offer potential for automating engineering design specifications but frequently generate performance claims lacking structural justification. This study investigates whether adversarial multi-agent system (MAS) architectures can enforce structural–functional balance by simulating professional engineering team deliberation. Three generative pipelines: zero-shot generation, iterative single-agent refinement, and adversarial MAS were evaluated across eleven consumer products spanning thermal appliances, motor systems, electromechanical assemblies, and structural mechanisms. Computational analysis employed the Axiomatic Realisability Index (ARI), measuring the ratio of structural provisions (materials, manufacturing) to behavioural requirements (performance targets, numerical constraints). MAS achieved substantially higher ARI across all products (mean 0.46 vs. 0.18, p<0.001, Cohen’s d=1.75, 100% directional consistency, +166% mean improvement). Critically, single-agent iterative refinement provided minimal benefit over baseline generation (mean ARI 0.18 vs. 0.20, p=0.48), with 50% of products showing degradation despite substantial document expansion (+ 77% word count, +68% technical density). Expert evaluation (N=5) revealed a perception challenge: while computational metrics identified MAS outputs as structurally superior, experts rated single-agent specifications higher in verifiability (4.8 vs. 3.8), correlating with numerical density rather than structural grounding. These findings demonstrate MAS effectiveness for constraint-heavy specification tasks while identifying deployment challenges requiring hybrid approaches addressing professional expectations for quantitative specificity. This study contributes a validated computational framework for assessing specification structural quality, systematic evidence that pipeline configuration, not simple iteration, determines output validity, and practical guidance for deploying AI-assisted specification tools in professional engineering design contexts.