Lessons from GitHub's Accessibility Agent Experiment: How AI Can Help Remove Digital Barriers

Introduction

GitHub has embraced the agent-based approach to streamline code creation and editing. As part of this shift, the company launched an experimental general-purpose accessibility agent aimed at two primary objectives: delivering just-in-time accessibility guidance to engineers via GitHub Copilot CLI and VS Code integration, and automatically detecting and fixing simple, objective accessibility issues before they reach production. The agent focuses on evaluating front-end changes, and early results are promising—over 3,535 pull requests have been reviewed with a 68% resolution rate.

Core Goals of the Accessibility Agent

Real-Time Accessibility Support

Engineers often need quick answers to accessibility questions while coding. By integrating directly with Copilot, the agent provides reliable, just-in-time responses that help developers make informed decisions without leaving their workflow. This reduces delays and encourages proactive accessibility practices.

Automated Remediation of Common Issues

The second goal is catching and fixing objective accessibility problems before deployment. The agent automatically examines pull requests that alter front-end code, flagging and in many cases patching issues. This proactive approach prevents barriers from reaching users who rely on assistive technologies.

Results and Key Findings

To date, the agent has analyzed 3,535 pull requests, achieving a 68% resolution rate. The most frequently identified issues fall into these five categories (in order of occurrence):

1. Making structure and relationships clear to assistive technologies
2. Providing clear and concise names for interactive controls
3. Ensuring users are aware of important announcements
4. Ensuring there are text alternatives for non-text content
5. Moving keyboard focus through pages and views in a logical order

Each resolved issue represents a friction point eliminated for people who use assistive tools, making GitHub more inclusive.

The Mindset Behind the Agent

Adopting the social model of disability, the team recognizes that barriers—and thus impairment—often arise from how environments are constructed. The agent is not a silver bullet; it cannot address every accessibility scenario. Instead, it augments human effort, helping colleagues remove the barriers that may be created when building GitHub's interfaces. Framing the agent as a supportive tool rather than a complete solution helped set realistic expectations and accelerated buy-in across teams.

Lessons Learned and Best Practices

Start Small, Focus on Objective Issues

The agent targeted simple, verifiable accessibility violations first. This allowed the team to demonstrate value quickly and build confidence before tackling more complex problems.

Integrate Seamlessly into Existing Workflows

By embedding accessibility checks within Copilot and pull request reviews, the agent became a natural part of development rather than an additional hurdle. Engineers could receive feedback without switching context.

Embrace the Social Model

Understanding that accessibility is about removing environmental barriers helped the team prioritize fixes that have the most impact on users. The agent's scope was defined to avoid overpromising, which in turn fostered trust.

Measure What Matters

Tracking resolution rates and issue types provided concrete data to guide improvements. The 68% resolution rate is a strong indicator of the agent's effectiveness, while the top issue categories highlight areas for further training.

Looking Ahead

The experiment continues, with plans to expand the agent's capabilities and refine its detection algorithms. GitHub hopes that sharing these findings will help other teams incorporate AI-driven accessibility tools into their own workflows. While the agent cannot replace human expertise, it serves as a powerful assistant in the journey toward more inclusive digital experiences.

For more background on LLMs and agents, see our related posts:Understanding LLMs, Designing Effective Agents.