An Agentic, Workflow-Native System for Generating Structured Accessibility Annotations in Design Tools

Abstract - Accessibility defects in modern software development frequently originate during the visual design stage, where dynamic interactions, focus management, and assistive technology expectations are insufficiently specified. Existing accessibility plugins within design environments predominantly function as static linters (e.g., color contrast checkers) or manual documentation kits, leaving critical behavioral accessibility guidance deferred to the engineering phase. This paper introduces a workflow-native, agentic system embedded within Figma and synchronized with developer workspaces (e.g., VS Code) that autonomously generates structured, screen-specific accessibility annotation reports. Guided by strict heuristic and WAI-ARIA Authoring Practices, the system utilizes a multi-model Large Language Model (LLM) engine to parse design artifacts and output deterministic Markdown documentation detailing keyboard interaction sequences, focus order, and screen reader announcements. Through an expert agreement analysis across 24 enterprise screens and a downstream quality assurance (QA) implementation study, we demonstrate that the system achieves substantial expert alignment (κ = 0.76) and reduces downstream accessibility defects by 32%. We argue that embedding AI-driven, documentation-first accessibility agents directly into the design-to-development pipeline effectively mitigates structural gaps in accessible product engineering. Key Words: Web Accessibility, Large Language Models, Human-Computer Interaction, Design Workflows, Assistive Technology, WAI-ARIA.