GitHub Pilots AI for Accessibility Automation

GitHub is piloting an AI agent designed to automate accessibility checks and fixes within pull requests, achieving a 68% resolution rate in early tests. The agent has already reviewed 3,535 pull requests, targeting structural issues, text alternatives, and keyboard focus problems by leveraging a robust dataset of past accessibility bugs.

The system relies on the non-deterministic nature of large language models to extrapolate patterns from historical data, enabling it to suggest fixes that often match human-level reasoning. To control costs and token usage, GitHub employs a sub-agent architecture that delegates specific subtasks — such as analyzing code diffs or generating patch suggestions — to smaller, specialized models, ensuring efficient processing even for complex accessibility challenges.

Early results suggest the agent can significantly reduce manual QA workload, though it still requires human review for edge cases. GitHub plans to expand the pilot based on feedback, aiming for broader integration into the developer workflow to make accessibility a seamless part of the code review process.