How GitHub's AI Agent is Automatically Fixing Accessibility Issues Before They Reach Users

GitHub is piloting an experimental general-purpose accessibility agent integrated with GitHub Copilot. This AI-driven tool aims to help engineers build more accessible interfaces by providing real-time guidance and automatically fixing common issues before code goes to production. In this Q&A, we explore how the agent works, its results so far, and the lessons learned from this initiative.

1. What is GitHub's new accessibility agent and why was it created?

GitHub's accessibility agent is an experimental tool built on large language models (LLMs) and agent technology. It integrates with both the GitHub Copilot CLI and the Copilot VS Code extension. The agent was created to address two pressing needs: first, to give engineers reliable, just-in-time answers to accessibility questions directly in their development environment; second, to automatically detect and fix simple, objective accessibility issues before code is shipped. The ultimate goal is to reduce barriers that would otherwise hinder users who rely on assistive technologies. By catching issues early, the agent helps teams build a more inclusive GitHub experience without requiring deep accessibility expertise from every developer.

2. What are the two primary goals of the accessibility agent?

The agent has two clear objectives. Goal one is to provide engineers with immediate, accurate accessibility guidance within the GitHub Copilot CLI and VS Code integration. This means developers can ask questions like "Does this button have an accessible name?" and get an answer without leaving their workflow. Goal two is to automatically evaluate code changes that affect front-end components and remediate straightforward issues before they reach production. For example, if a new element lacks proper ARIA labels or has incorrect keyboard focus order, the agent can flag it and even suggest or apply fixes. This dual approach not only educates developers but also prevents many common accessibility errors from ever affecting users.

3. What results has the agent achieved so far in terms of pull request reviews?

Since its pilot began, the accessibility agent has reviewed 3,535 pull requests with a 68% resolution rate. That means over two-thirds of the issues the agent identified were successfully addressed before merging. While the agent doesn't catch every possible accessibility problem, it focuses on objective, testable violations—things that can be automatically detected and fixed. The high resolution rate shows that when developers receive clear, actionable feedback, they act on it. This has significantly reduced the number of common barriers that would have otherwise reached production, directly benefiting users of assistive technology.

4. What are the top five accessibility issues the agent catches most frequently?

Based on the frequency of detections, the agent's top five issue types are:

• Structure and relationships: Ensuring that landmarks, headings, and table structures are correctly conveyed to assistive technologies.
• Clear names for interactive controls: Buttons, links, and form inputs must have accessible names that describe their purpose.
• Important announcements: Dynamic content changes (like status updates) need to be announced via live regions.
• Text alternatives: Non-text content (images, icons) must have proper alt text or aria-labels.
• Logical keyboard focus order: Focus should move through interactive elements in a meaningful sequence.

These five categories represent the most common friction points that can block assistive technology users from navigating GitHub effectively.

5. How does the social model of disability influence the agent's design?

GitHub's team approaches accessibility through the lens of the social model of disability, which holds that barriers—and therefore impairment—are created by how environments (including digital ones) are built. Rather than trying to "solve" accessibility in isolation, the agent is designed to augment human effort. It helps developers remove the barriers they might unknowingly introduce while constructing GitHub's user interfaces. This mindset shifts responsibility from individual users to the product team and the tools they use. The agent is not a silver bullet; it's a practical assistant that makes it easier for engineers to uphold inclusive design practices, reducing the gap between intention and outcome.

6. What important limitations does GitHub acknowledge about this agent?

GitHub explicitly states that the accessibility agent is not a silver bullet. It cannot address every hypothetical accessibility scenario, especially those that require human judgment, context, or user testing. The agent is scoped to catch simple, objective issues that are reliably testable. This self-awareness was key to getting the experiment launched quickly and gaining buy-in from teams. By setting clear boundaries on what the agent can and cannot do, GitHub avoided overpromising and focused on delivering real value where automation works best. They emphasize that the agent is meant to complement—not replace—the expertise of accessibility specialists, manual testing, and user feedback.