Automated accessibility reviews now block 68% of detected issues in pull requests
GitHub has launched an experimental general-purpose accessibility agent that automatically reviews code changes for barriers affecting users of assistive technology. The AI-driven system has already scanned 3,535 pull requests and resolved 68% of detected issues before they reach production.
"This isn't about replacing human judgment—it's about catching the objective, repeatable issues that slip through," said a GitHub accessibility engineering lead. "Engineers get just-in-time suggestions while they code, and automatic remediation catches the rest."
Top five accessibility issues automatically removed
The agent identifies and fixes common barriers in front-end code. The most frequent problems, in order, are:
- Unclear structure and relationships for assistive technologies
- Missing or confusing names for interactive controls
- Unannounced important updates for screen reader users
- Missing text alternatives for non-text content
- Illogical keyboard focus order through pages
Each fix eliminates friction that would have blocked GitHub usage for people relying on assistive tech.
Background: The rise of LLM-powered agents at GitHub
AI agents have become central to GitHub's development workflow. The company already uses agent-based code creation and editing across many initiatives, and the accessibility agent extends that approach to inclusion.
"Agents are powerful, but they're not magic," the lead explained. "We grounded this experiment in the social model of disability—access barriers are created by how environments are built. This agent doesn't solve accessibility alone; it augments human effort."
GitHub has published resources on LLMs and agent frameworks to help teams build similar tools. Key guides include A guide to deciding what AI model to use and Multi-agent workflows: How to engineer ones that don't fail.
What this means for web accessibility
Early results suggest that even a general-purpose agent can catch a meaningful percentage of objective accessibility errors. The 68% resolution rate covers simple, machine-detectable issues—not nuanced design decisions.
"We're not aiming for a silver bullet. By scoping the agent's responsibility, we got faster buy-in and a working tool in production," said the lead. "The goal is to remove friction at scale, so engineers can focus on the harder, subjective problems."
For the millions of users who depend on assistive technology, automated catching of barriers before deployment means fewer broken experiences. GitHub plans to share detailed lessons to help other teams replicate the approach.
— Updated with latest PR review data from GitHub's internal experiment.