Introduction
GitHub recently piloted a general-purpose accessibility agent integrated with GitHub Copilot. Its aim is to help engineers catch and fix common accessibility issues before they reach production, while also providing on-demand guidance. After reviewing 3,535 pull requests and achieving a 68% resolution rate, the team shared valuable insights about the agent's design, mindset, and impact. Below, we answer key questions about this experiment.
What is GitHub’s accessibility agent and what does it do?
The accessibility agent is an experimental AI tool built for GitHub Copilot, working within both the CLI and VS Code environments. Its primary function is twofold: first, it provides engineers with reliable, just-in-time answers to accessibility questions—helping them make informed decisions during development. Second, it automatically detects and remediates simple, objective accessibility issues in front-end code before that code reaches production. The agent is not meant to solve every possible scenario; instead, it focuses on clear-cut, repetitive problems that can be identified programmatically. By catching these issues early, the agent reduces the burden on developers and helps prevent barriers for users of assistive technology.
What are the two main goals of the accessibility agent?
The agent was designed to achieve two specific goals. The first is to provide just-in-time accessibility answers directly within the Copilot workflow—so engineers don’t have to leave their code to look up guidelines. The second is to catch and automatically fix common, objective accessibility issues before they go to production. For the second goal, the agent automatically evaluates any change that modifies GitHub’s front-end code. These goals complement each other: the Q&A feature educates developers, while the automated remediation reduces the number of issues that slip through. Together, they aim to improve the accessibility of GitHub’s user interface incrementally, without overwhelming the team.
How effective has the agent been so far?
As reported, the agent reviewed 3,535 pull requests and achieved a 68% resolution rate. This means that in over two-thirds of cases, the identified accessibility issues were either automatically fixed or resolved without further human intervention. This is a strong indicator that the agent is successfully catching problems that might otherwise reach production. The top five issue types it finds (see the next section) represent real friction for assistive technology users, and the agent’s ability to address them automatically has already reduced barriers on GitHub. While not a perfect tool—some issues require human judgment—the agent has proven to be a valuable addition to the development workflow.
What are the top five types of accessibility issues the agent catches?
In order of occurrence, the five most common issues are:
- Structure and relationships: Making sure headings, landmarks, and other semantic structures are clear for assistive technologies.
- Interactive control names: Providing clear, concise names for buttons, links, and form controls so screen readers can announce them properly.
- Important announcements: Ensuring dynamic content changes are announced to users of assistive technology.
- Text alternatives: Making sure all non-text content (images, icons) has a meaningful text alternative.
- Logical keyboard focus: Moving keyboard focus through pages and views in a predictable order.
These issues represent common, objective barriers that the agent can automatically detect and often remediate. By removing them early, the agent helps create a more inclusive experience for all GitHub users.
What mindset guided the development of the accessibility agent?
The team adopted the social model of disability, which teaches that barriers—and therefore impairment—are created by the environment, not by the individual. For digital experiences, that means accessibility issues arise from how code is written, not from the user. With this mindset, the accessibility agent is not seen as a “silver bullet” that can solve everything, but rather as a tool to augment human effort. The team focused on removing barriers that are commonly created during the construction of GitHub’s user interfaces. Setting this clear scope helped speed up the experiment’s launch and increased buy-in across the organization. The agent works best when paired with human judgment, not as a replacement for it.
What key lessons were learned during this experiment?
Several important lessons emerged. First, keeping a narrow scope is critical—the agent doesn’t attempt to solve every accessibility issue; it targets objective, frequent problems. This focus allowed the team to launch faster and demonstrate value early. Second, clear communication about what the agent can and cannot do helped manage expectations and build trust. Third, integrating the agent early in the development pipeline (before code is production) maximized its impact. Finally, the agent works best as a complement to human expertise—developers still need to handle complex, subjective accessibility decisions. These lessons can guide other teams looking to build similar tools for their own workflows.