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

GitHub is piloting an experimental general-purpose accessibility agent integrated with GitHub Copilot that reviews pull requests for accessibility issues and auto-remediates simple ones. After reviewing 3,535 PRs with a 68% resolution rate, the team shares key lessons: the agent uses a two-sub-agent architecture (reviewer + implementer) with sandboxed communication via template schemas, executes instructions in linear order to improve accuracy, and avoids high-risk patterns like drag-and-drop and data grids. Critical insights include the importance of pre-existing manually audited accessibility issue corpora to counteract LLMs' bias toward inaccessible code, using complexity scoring to route complex cases to human experts, and understanding that ~36% of WCAG A/AA criteria cannot be automatically detected. The agent is positioned as an augmentation tool, not a replacement for human accessibility expertise.