Triple Loop AI Engineering — A Three-Layer Iterative Design Theory for AI Agent Systems

This paper proposes Triple Loop AI Engineering, a three-layer iterative design theory for AI agent systems that addresses the fundamental limitation of single-loop architectures prevalent in current agentic AI design. Drawing on Argyris’s double-loop learning theory (1977) as a theoretical foundation, and on the “triple-loop” extension developed in organizational learning theory (Swieringa & Wierdsma, 1992; Tosey et al., 2012), we introduce the Design Diagnostic Protocol (DDP)—a structured five-category diagnostic framework for AI-specific failure modes—and extend it with a Meta Loop (Loop 3) that questions the problem definition itself when lower-level loops fail to converge.We report an exploratory empirical validation across four controlled experiments conducted within a real business workflow (new product planning for consumer brands) at a Japanese fabless manufacturer. Key findings: (1) Triple Loop increased the number of recommended products in STEP2 from 0 to 3 compared to a Single Loop control and improved the mean STEP2 composite score by +0.33 on average (mean across all candidates in each condition) (E1 vs. E2); (2) across N=5 DDP events spanning two brands and four workflow steps (E1, E3), every DDP diagnosis matched the author’s post-hoc root-cause assessment—we emphasize this was a single author’s judgment without an independent evaluator, so the result is illustrative of the five-category framework’s applicability rather than a validated accuracy rate; (3) Loop 3 successfully triggered under D_max=2 conditions in a constructed scenario in which brand, target segment, and success criteria were intentionally withheld from the input, reformulated the problem from “select a product” to “first establish brand × target × KPI,” and achieved convergence in the first subsequent Loop 1 iteration without DDP intervention (E4). A further observation ofinterest is cross-brand learning: a threshold correction diagnosed in one brand (E1) enabled first-iteration, DDP-free convergence when applied to a second brand (E3). These results provide preliminary evidence for the practical feasibility of multi-layer AI loop engineering in complex, real-world business tasks.