A11yAuditor: An Agentic Workflow for Focused Reasoning and Precise Context Retrieval in Automated Web Accessibility Auditing

Abstract

Large language model-based agents are increasingly used for automated web accessibility auditing by invoking existing hard-coded checkers as tools. However, current approaches suffer from two fundamental limitations: (1) a broad task scope requires a single agent to audit multiple accessibility criteria simultaneously, causing reasoning interference; and (2) noisy observations introduced by tool outputs force the agent to reason over large amounts of criterion-irrelevant context, degrading audit performance. To address these limitations, we propose A11yAuditor, a novel agentic workflow consisting of two complementary components: (i) per-criterion audit decomposes accessibility auditing into dedicated criterion-specific agents, transforming broad task scopes into focused reasoning tasks; and (ii) agentic search persists tool outputs in a virtual filesystem and enables each agent to actively retrieve only criterion-relevant evidence, enabling precise context retrieval. Together, these components reduce reasoning interference and context noise during agent execution. We evaluate both A11yAuditor and a representative tool-using agent baseline against real-world accessibility audit reports produced by certified human experts as ground truth. Experimental results show that A11yAuditor significantly improves accessibility audit performance, achieving higher precision, recall, and F1 through focused task scopes and precise context retrieval. Ablation studies reveal that per-criterion audit and agentic search contribute synergistically, while scalability experiments demonstrate that the proposed workflow remains robust as the number of accessibility criteria increases.

Related Projects

• 전문가 맥락 기반 Multi-Agent AI 기반 디지털 접근성 자동 진단 및 UI 자동 생성 기술 개발

  2026.07 — 2027.03

• 전문가·사용자 생성형 AI 기반 UX 접근성 평가 자동화를 위한 GUI 에이전트 기술 개발 및 검증

  2025.09 — 2026.08