Lessons from Building a General-Purpose Accessibility Agent

GitHub is experimenting with a general-purpose accessibility agent in GitHub Copilot CLI and Copilot VS Code integration that instantly answers accessibility questions and automatically fixes simple, objective accessibility issues in frontend changes. So far it has reviewed 3,535 PRs and resolved 68% of them. The top issues caught repeatedly were conveying structure and relationships, naming interactive controls, status messages, alt text for non-text content, and keyboard focus order.

Accessibility is an issue of an environment created together by code, design, and copywriting. The team treats this not as something to be solved alone, but as a way to augment colleagues' work and remove barriers. The EAA is already in effect, and the U.S. ADA Title II will make WCAG 2.1 AA the legal completion standard in April 2027.

The training material was GitHub's long-accumulated manual issue management system. Issue templates, reproduction steps, severity/service area/WCAG success criteria metadata, related PRs, and acceptance criteria are all gathered in one repository, forming a structured corpus from which the agent can infer relevant code and copy. Without this accumulation, LLMs easily reproduce accessibility anti-patterns.

To reduce token consumption and incorrect answers, the architecture is split into a parent orchestrator and two dedicated subagents. Instead of a large group of subagents, there are only a Reviewer and an Implementer.

• Reviewer: audits code, investigates WCAG, detects escalation signals, produces structured results
• Implementer: defaults to code-change mode, switches to guidance-only mode when complexity is high
• Parent: routes requests, explores code and skills, scores complexity, validates, manages recheck loops and escalation gates

They don't exchange content directly with each other; the parent only receives templated output and validates it. This separation is designed to manage token consumption, response speed, and operating costs.