GitHubが汎用アクセシビリティエージェントを開発、その過程で得た知見 Building a general-purpose accessibility agent—and what we learned in the process

GitHub has built a general-purpose accessibility agent that automatically discovers and proposes fixes for accessibility issues in websites and applications. Built on top of Copilot's underlying technology, the agent aims to move beyond code completion toward understanding UI structure and semantic markup, supporting developers in delivering experiences that work for screen reader users and others who rely on assistive technology.

Accessibility remains a persistent challenge in software development. International standards such as WCAG provide clear targets, but in practice accessibility work is often deferred or handled superficially. Existing static analyzers like axe-core and Lighthouse can catch mechanically detectable issues such as missing alt text or insufficient color contrast, but they tend to struggle with context-dependent judgments like whether a focus order is logical or whether ARIA attributes are semantically appropriate. GitHub's agent appears to target precisely this gray zone, leveraging LLM-based contextual reasoning to evaluate intent rather than just syntax.

A recurring theme in the lessons shared is that building a useful general-purpose agent is less about the model itself and more about the surrounding scaffolding. The team emphasizes giving the agent access to the right domain-specific tools—accessibility scanners, browser automation for rendering and interaction, and code editing capabilities—and orchestrating them in a way the model can reason over. Pure LLM output is prone to hallucinated fixes or overcorrection, so combining probabilistic reasoning with deterministic verification steps emerged as critical to quality. This mirrors the broader industry consensus, echoed by Anthropic and OpenAI, that tool-using agents outperform monolithic prompts on practical tasks.

Another insight is the importance of iteration loops. Rather than trying to fix everything in one pass, the agent benefits from a cycle of detect, patch, re-test, and refine. This pattern is becoming standard for coding agents in general, with frameworks like SWE-agent and Cursor's background agents adopting similar architectures.

The accessibility tooling space is becoming increasingly crowded. Microsoft's Accessibility Insights, Deque's Axe DevTools, and various startup offerings are all pushing automation further, and the question of how much can be safely automated remains open. Automated PRs that fix accessibility issues can dramatically reduce reviewer burden, but a poorly judged fix—say, an ARIA label that misrepresents intent—can actively harm users who depend on assistive tech. Human-in-the-loop review is therefore likely to remain essential, at least for the foreseeable future.

For GitHub, the agent also serves as a testbed for general agent-building practice. Many of the patterns surfaced here—tool integration, verification loops, scoped autonomy—are likely to inform how Copilot's broader agentic capabilities evolve across other domains beyond accessibility.