Lessons from GitHub's Accessibility Agent: An Experimental Q&A

GitHub recently experimented with a general-purpose accessibility agent designed to help engineers build more inclusive software. This Q&A explores the agent's goals, results, and key takeaways from the initiative.

What was the main purpose of GitHub's accessibility agent?

The accessibility agent was created to serve two primary functions. First, it provided engineers with just-in-time answers to accessibility questions directly within GitHub Copilot CLI and VS Code integration. Second, it automatically identified and fixed simple, objective accessibility issues before they reached production. By integrating this agent into the development workflow, GitHub aimed to remove common barriers that hinder users of assistive technologies, such as screen readers or keyboard navigation tools. The agent reviewed code changes that modified front-end components, ensuring that accessibility was considered from the start rather than as an afterthought. This proactive approach helped reduce friction for both developers and end users, making GitHub more inclusive without adding extra burden to engineering teams.

How successful was the agent in reviewing pull requests?

Over the course of the experiment, the accessibility agent reviewed 3,535 pull requests with a 68% resolution rate. This means that in more than two-thirds of the cases, the issues identified were automatically fixed or addressed before merging. The top five issue types, ranked by frequency, included: making structure and relationships clear to assistive technologies, providing clear and concise names for interactive controls, ensuring users are aware of important announcements, adding text alternatives for non-text content, and maintaining a logical keyboard focus order. Each of these represents a real barrier that, if left unresolved, would have made GitHub harder to use for people relying on assistive technology. The agent's ability to catch these issues automatically saved engineers significant time and improved the overall accessibility of the platform.

What was the mindset behind the accessibility agent?

The team adopted a social model of disability approach, recognizing that accessibility barriers are created by how environments—including digital interfaces—are built. Rather than aiming to “solve” accessibility in isolation, the agent was designed to augment the efforts of GitHub engineers. It was never intended to be a silver bullet that automatically addresses every possible scenario. By clearly defining the agent's scope and limitations, the team set realistic expectations and secured broader buy-in. This mindset helped accelerate the experiment's launch and fostered a culture where accessibility is a shared responsibility, not just a final checklist item. The goal was to remove the most common, objective obstacles efficiently, while still empowering human judgment for more complex or subjective accessibility challenges.

What were the top five issue types the agent caught most often?

Based on the experiment, the most frequent issues detected by the agent were:

• Structure and relationships – Ensuring that headings, lists, and landmarks are properly coded so screen readers can interpret page layout.
• Clear control names – Interactive elements like buttons and links must have descriptive labels for voice commands and screen readers.
• Important announcements – Users must be informed of dynamic content changes, such as errors or confirmations, via ARIA live regions.
• Text alternatives – All non-text content (images, icons, charts) needs alt text or equivalent descriptions.
• Keyboard focus order – The tab order must follow a logical sequence so that keyboard-only users can navigate without confusion.

Automatically resolving these issues prevented significant friction for people using assistive technologies, contributing to a more inclusive GitHub experience.

What lessons did GitHub learn from this experiment?

One key lesson was that setting clear, narrow goals for the agent helped it deliver value quickly. Instead of trying to fix every accessibility concern, the team focused on objective, repeatable issues. This made the agent reliable and earned engineers' trust. Another insight was the importance of integrating the agent into existing tools like Copilot—so developers didn't need to change their workflow. The experiment also emphasized that automated tools work best when paired with human expertise; the agent handled the routine fixes, while engineers addressed nuanced decisions. Finally, sharing results transparently across teams increased enthusiasm for broader accessibility initiatives. The 68% resolution rate demonstrated that even partial automation can have a meaningful impact on reducing barriers.

How does this agent relate to broader accessibility efforts at GitHub?

The accessibility agent is part of a larger commitment to make GitHub usable by everyone, regardless of ability. While the agent tackles simple, objective issues, it complements other accessibility work like design reviews, manual testing, and developer training. GitHub also invests in community resources and documentation to spread accessibility best practices. The experiment showed that automation can scale certain aspects of accessibility enforcement without replacing human judgment. By reducing the number of common errors, the agent frees up accessibility specialists to focus on more complex challenges. Ultimately, the goal is to embed accessibility into every stage of development—from planning to deployment—so that it becomes a natural part of the engineering culture rather than a separate compliance exercise.