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

2026-05-15 · Source: The GitHub Blog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Intermediate, long

Summary

GitHub is piloting an experimental general-purpose accessibility agent designed to provide engineers with just-in-time answers to accessibility questions within GitHub Copilot CLI and VS Code, and to automatically remediate simple, objective accessibility issues before production. To date, the agent has reviewed 3,535 pull requests, achieving a 68% resolution rate for common issues like clear structure, control names, and text alternatives. The agent operates on a sub-agent architecture, separating passive review/research from active implementation, and uses a linear execution order with template schemas for consistency and efficiency. It acknowledges limitations by evaluating code complexity, avoiding high-risk patterns, and incorporating anti-gaming instructions to prevent over-generation. This initiative augments human efforts, recognizing that approximately 36% of WCAG A and AA Success Criteria still require manual evaluation.

Key takeaway

For AI Engineers developing specialized LLM agents, you should prioritize augmenting human expertise rather than attempting full automation. Focus on training agents with manually audited, structured data to counteract LLM biases towards antipatterns. Implement a sub-agent architecture with linear instruction execution and template schemas to manage token consumption and ensure accuracy. Critically, integrate mechanisms to identify code complexity and high-risk patterns, directing human intervention where agents fall short to prevent costly reworks and reputational damage.

Key insights

An accessibility agent augments human efforts by automating issue remediation and providing guidance, but requires careful design and human oversight.

Principles

• Accessibility agents augment, not replace, human expertise.
• Structured, manually curated data improves agent effectiveness.
• Sub-agent architectures enhance efficiency and accuracy.

Method

Implement a sub-agent architecture with sandboxed reviewer and implementer agents. Enforce linear instruction execution using template schemas for structured output, and integrate complexity scoring and high-risk pattern avoidance.

In practice

• Invest in manual accessibility issue cataloging.
• Use a sub-agent model for complex LLM tasks.
• Define template schemas for inter-agent communication.

Topics

• LLM Agents
• Accessibility
• Sub-agent Architecture
• Automated Remediation
• WCAG Compliance
• GitHub Copilot

Best for: AI Engineer, Software Engineer, Director of AI/ML

Related on AIssential

• Step-by-Step Guide to Building AI Agents Using LLMs — LLM-powered AI agents enable advanced automation by understanding language, reasoning, and interacting with systems autonomously.
• To LLM, or Not to LLM: How Designers and Developers Navigate LLMs as Tools or Teammates — LLM adoption is a sociotechnical positioning problem, not merely a technical evaluation of capability.
• Teaching Machines to Code: A Social Scientist’s Guide to LLM-Based Information Extraction — Multi-agent LLM pipelines excel at complex information extraction by mimicking human collaborative coding processes.
• How to Make a Coding Agent a General Purpose Agent - Harrison Chase — Coding agents, with robust harnesses and secure tool runtimes, form the foundation for scalable, general-purpose enterprise AI agents.
• Anatomy of Agent SKILLS — Agent "skills" prevent context bloat by progressively disclosing information, improving efficiency and modularity.
• The Amazing Agent Race: Strong Tool Users, Weak Navigators — LLM agents struggle with navigation in complex, non-linear tasks more than with tool execution.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The GitHub Blog.