Large language models offer a scalable means of supporting expert judgment elicitation in analytic hierarchy process decision making, but their pairwise judgments may still violate consistency requirements in comparison-intensive hierarchical tasks. This study develops a multi-agent iterative elicitation framework that treats inconsistency as an elicitation-stage convergence problem rather than a post hoc matrix repair task. To operationalize this framework, a stage-adaptive focus-pair revision mechanism is proposed to select revisable comparisons from inconsistent matrices, provide rule-based soft hints, and adjust revision priority and scope according to the current inconsistency regime. A CubeSat design-and-development risk assessment task is used for validation. Across twenty independent runs, all inconsistent revision trajectories converged monotonically to acceptable consistency, and the treated outputs preserved a decision-plausible high-priority risk layer. Compared with a Xu–Wei direct correction baseline, SA-FPR achieved comparable top-priority stability with substantially fewer pairwise judgment modifications, while retaining traceable focus-pair revisions and expert-side rationales.