The three building blocks

The Ability profile agent

It keeps a record of what you ask for, e.g., larger text or voice control. It doesn’t record a diagnosis or a condition. Apps can read only what you allow, and you can see the record, correct it and withdraw permission whenever you like.

The same profile works in any app or on any device you allow, e.g., a website, a phone app or a headset. It can stay inside one app if you don’t want an outside service, or it can run as a separate service that the apps you allow can reach.

You don’t have to fill in a long survey to set up a profile. People in our early sessions told us they don’t like having to pick a category that describes them. So the toolkit starts from what you say is hard, in your own words. We’re also testing whether it can learn from how you use an app and how you correct its changes, so that you don’t have to describe yourself at all.

The Adapter agent

Adapters are small pieces of code, and each one changes one thing in an app, e.g., making text larger, simplifying a layout, adding captions or turning a chart into a table. The toolkit includes about 50 of them.

Adapters combine into skills. A skill is a short recipe for a need, such as larger text with less motion and fewer distractions. If you ask for something no skill covers yet, the Adapter agent can put together a new skill from the existing adapters. Nothing is saved until you approve it.

This is different from an accessibility overlay. A website’s owner adds an overlay, often to pass a compliance check, and every visitor gets the same changes. With the toolkit, you choose each change for yourself.

The Verification agent

It checks that a change did what you asked without breaking anything that worked. When an AI agent is acting for you, it stops the agent before decisions that are yours to make.

In one demonstration, an AI agent booking a hotel for a screen reader user stopped three times: when “Stanford” matched five different places, when a room would have put one guest on a sofa bed, and when taxes at checkout pushed the price over budget. Each time, the person decided what to do.

Today the Verification agent can confirm that a change was made and didn’t break the page. It can’t yet tell whether the change made the page more accessible for a particular person. That needs an agreed way to measure accessibility, and there isn’t one yet.

You stay in control

  • Your profile records what you ask for, not a condition. Nobody has to say they’re disabled to use it.
  • Apps read only what you allow.
  • The toolkit suggests each change, and you approve it. A website can’t force a change on you.
  • An AI agent pauses at decisions that are yours to make, such as paying.
  • You can undo a change, and you can see when an agent is acting for you and stop it.

Beyond the browser

Nothing in a profile, a skill or the building blocks assumes a web page. We started with the browser because that’s where most reading happens today, but the same profile is meant to follow a person onto other devices. A voice guide for smart glasses already uses it to set the pace and wording of its instructions. The toolkit can also turn a profile into settings for extended reality (XR) headsets, e.g., comfortable text sizes and captions that stay fixed in place.