Building a general-purpose accessibility agent—and what we learned in the process - The GitHub Blog
Eric Bailey·@ericwbailey
May 15, 2026
| 11 minutes
Share:
It is an understatement to say agents have become a popular way of working with code. GitHub has adopted agent-based code creation and editing for many of its initiatives, including piloting an agent to help with our commitment to accessibility.
GitHub is currently piloting an experimental general-purpose accessibility agent to achieve two main goals:
Providing engineers with reliable, just-in-time answers to accessibility questions in the GitHub Copilot CLI and the Copilot VS Code integration.
Catching and automatically remediating simple, objective accessibility issues before they go to production.
For purpose number two, the accessibility agent is set to automatically evaluate changes that modify our front-end code.
To date, the agent has reviewed 3,535 pull requests, with a 68% resolution rate. In order of occurrence, the top five issue types center around:
Making structure and relationships clear to assistive technologies
Providing clear and concise names for interactive controls
Ensuring users are aware of important announcements
Ensuring there are text alternatives for non-text content
Moving keyboard focus through pages and views in a logical order
Each of these issue types represents friction and barriers automatically removed that would have otherwise inhibited use of GitHub for people who use and rely on assistive technology. Here’s a screenshot of it in action:
Interested? We’ll be outlining successes and lessons learned with this experiment, with the hopes that it can help with other teams’ accessibility journeys.
Mindset
The social model of disability teaches us that access barriers—and consequently impairment—can be created because of how an environment is built. The same thinking applies to digital experiences.
With the accessibility agent, we are not attempting to “solve” accessibility in isolation. We are instead attempting to augment our peers’ efforts, to better help them remove the barriers that may be created as a result of how we construct GitHub’s user interfaces.
The accessibility agent is not a “silver bullet” that can automatically address every hypothetical scenario. Understanding, honoring, and socializing this better helps set the agent’s scope of responsibility. This sped up the experiment’s launch, leading to more buy-in for the effort.
Past efforts
The European Accessibility Act is now in effect. Title II of the Americans with Disabilities Act is set to establish meeting WCAG 2.1 AA as the legal definition of done in April of 2027. LLM agents can read and take action on the accessibility tree.
To say it plainly: Organizations will be at a disadvantage if they have not already invested in manually identifying and remediating accessibility issues. There are many reasons for this, including building an accessibility agent.
To that point, GitHub has a mature system in place for logging accessibility issues, as well as verifying fixes to issues are working as intended. This includes:
A structured template for reporting problems
Steps to reproduce the issue
A rich layer of metadata about the issue’s severity level, service area, and applicable WCAG success criterion
Crosslinks to the Pull Request that addressed the issue
Acceptance criteria
In addition, all the issues are centralized to a single repository. While this issue-logging effort predated the explosion in popularity of LLM tooling, its highly consistent and structured nature made it an ideal corpus of content for the accessibility agent to reference.
Because of this, we instructed the agent to investigate these issues to see if there are related code and language snippets it can extrapolate from. This is one area where the non-deterministic “fuzzy matching” behavior of LLMs acts as an asset rather than a possible liability.
Old gold
Much like with any other specialized domain area, vague instructions in a skill file won’t cut it. Telling an LLM to “use accessibility best practices” with a short list of examples won’t work well.
When generating code, LLMs have an unfortunate bias towards producing accessibility antipatterns since every major LLM currently available is trained on decades of inaccessible code.
To counteract this, the agent needs better content to draw from.
So, I enthusiastically recommend investing in manually cataloging and remediating accessibility issues. After some progress, this data can be incorporated into the agent.
The issues and their corresponding pull requests provide highly contextual examples for the LMM to reference, written using the conventions set up by the organization it is deployed in. This collection of issues and code is by far one of the strongest assets the agent draws from.
Efficient token consumption
Accessibility is a holistic concern, intersecting with code, design, copywriting, and numerous other disciplines involved with creating user interfaces.
A lot of accessibility work is also highly contextual, meaning that someone typically needs the full working picture of a problem before they’re able to give the appropriate advice for what to do.
Because of these two factors, a general-purpose accessibility agent can consume a ton of tokens when it performs work. This has three negative outcomes:
An increased amount of unreliable output
Slower response times
Increased operational costs
It’s important to be diligent when structuring the agent. Here’s how we went about doing just that.
Use sub-agents
The accessibility agent started as a single monolithic agent, but quickly grew past the limitations of this approach. Because of this, we evolved it to use a sub-agent architecture.
A lot of guides recommend creating a large suite of sub-agents, each with its own specific area of responsibility. Here, the sub-agents are executed in parallel, with the main agent reconciling their output.
Surprisingly, this approach worked against us for the accessibility agent. Instead, we wound up using two dedicated sub-agents:
The first sub-agent acts as a passive reviewer and researcher.
The second sub-agent acts as an active implementer.
The two sub-agents are sandboxed and cannot directly pass content to each other. Instead, they generate a structured, templatized output. This output is then served to the parent orchestrating accessibility agent to consume, validate, and route.
There are a few reasons for this approach:
Escalation checkpoints. The reviewer checks for areas where human intervention will likely be needed. This includes multiple high-severity WCAG failures, as well as a list of patterns that are known to be difficult to make accessible.
Complexity-based behavior. The agent is instructed to operate in a specialized guidance-only mode if the underlying code is deemed too complicated. Here, the parent accessibility agent acts as an arbiter, while the reviewer agent is “opinionless” and just reports the findings as instructed.
Filtering. The reviewer presents everything it finds. The parent accessibility agent then utilizes resources and skills to determine what is relevant to the request. The reviewer passing all its findings to the implementer would be costly and potentially set it on irrelevant and counter-productive tasks.
Traceability. Direct communication between sub-agents would remove the ability to create and review an audit trail of user and agent decisions. This is important given the agent’s instruction around complex patterns, as well as the highly contextual nature of accessibility work.
Execute instructions in a linear order
In addition to being a holistic concern, effective digital accessibility work also demands a methodical, detail-oriented approach.
The concern of using sub-agents to increase the speed of the LLM’s reply is counterbalanced by our need for its results to be accurate. We found that compelling the agent to execute its sub-agent instructions in a fixed order was key.
We first establish a parent set of ordered phases. Each phase itself contains child ordered steps of instructions, which are accompanied by relevant resources and skills:
The interesting bit about this linear order is that it mirrors how I would personally approach performing auditing, remediating, and reporting duties.
Use a template schema pass around sub-agent content
The entire operation of the sandboxed sub-agents is built around template schema files. These files create consistency that is vital to keeping the agent focused and on track.
The two schema templates are:
Reviewer template schema: This focuses on what to audit, and how to find applicable information about it.
Implementer template schema: This focuses on what to fix and how to fix it.
Without the schema files in place, the agents would all attempt to arbitrarily communicate with each other. This would create increased token expenditure, undesirable hallucinations, unnecessary code changes, and difficult-to-impossible behavior for agent auditing purposes.
Acknowledging limitations
Another vital aspect of creating the accessibility agent is understanding areas where agents can fall short.
As the agent is not a turnkey “solution” for accessibility, we want to avoid situations where the agent’s output in error may not be sufficiently interrogated by the human using it. This is especially relevant when someone is not well-versed in digital accessibility considerations and practices.
Here’s what