GitHub Launches an AI Agent for Accessibility Error Detection and Fixing

GitHub is experimenting with an AI agent that checks code accessibility and automatically fixes common errors that prevent users of assistive technologies, like screen readers, from using the platform effectively.

How the Agent Works

At first, the team tried a monolithic architecture—one large agent for everything. But this required too many tokens and had a high probability of hallucinations. So GitHub redesigned the system into a two-tier architecture.

At the top level, a main orchestrator works. It routes requests, validates results, and manages escalation. At the lower level, two specialized agents work in sequence: first a reviewer checks code in read-only mode and finds problems, then an implementer either generates fixes or just provides suggestions.

The agents don't communicate directly—they only communicate through rigid, predefined templates. This reduces token consumption and almost completely eliminates hallucinations. Each step is performed in a strict, fixed order—as if the fixes were being made by a human accessibility specialist.

What Problems It Fixes

The agent catches five main categories of errors:

• Incorrect semantic markup — when HTML tags don't reflect the meaning and structure of the content
• Poor control element names — buttons and input fields without clear `<label>` or aria attributes
• Missing status announcements — when the system doesn't announce state changes (for example, validation errors)
• No text alternatives for images — missing `alt` attribute
• Confusing logical navigation — when keyboard focus moves unpredictably or incorrectly

During the pilot, the agent processed 3,535 pull requests and fixed issues in 68% of cases. For users who rely on screen readers and other assistive technologies, this makes a real difference in the usability of GitHub.

Challenges and Safeguards

GitHub encountered an interesting paradox: LLM models are trained on decades of historical code, including many examples of poor accessibility. So the models themselves often generate the same accessibility antipatterns that need to be fixed. For example, they can forget about `<label>` for an input or suggest incorrect aria structure.

The team added several safeguards. First, the agent doesn't touch complex patterns at all—drag-and-drop interfaces, tree view, data grid—it sends those for human review because they require deep expertise. Second, there's a script that analyzes code complexity; if it's above the threshold, the agent switches to guidance-only mode instead of trying to generate code independently.

It also turned out that historical data is more critical than general instructions. GitHub used its own database of 3,500 manually logged and corrected errors accumulated over the years. These real examples from their own code helped the agent learn best and generate relevant fixes.

What This Means

AI doesn't solve accessibility problems alone. But it significantly speeds up work and helps catch errors at an early stage. GitHub plans to open-source the agent so other projects and companies can apply it to their code. This could become a standard for accessibility checking in CI/CD pipelines.