Building a general-purpose accessibility agent—and what we learned in the process

GitHub has been experimenting with an accessibility agent integrated into its Copilot tools, aiming to provide just-in-time answers to accessibility questions and automatically remediate simple accessibility issues in its code. This agent has reviewed over 3,500 pull requests, achieving a 68% resolution rate for issues like clarifying structures for assistive technologies and ensuring text alternatives for non-text content. The agent operates using a sub-agent architecture, with one sub-agent acting as a passive reviewer and the other as an active implementer. It is designed to work around the limitations of LLMs, recognizing the nuanced nature of accessibility work and the importance of manual intervention for complex issues. GitHub's structured approach and focus on learning from the agent's output aim to ensure continuous improvement and integration of accessibility best practices, with the potential for future open-sourcing to benefit wider accessibility efforts in open-source software.