Room 302: Good day, everyone. This is a breakout session at TPAC 2025 from the Accessible Platform Architectures Working Group. We're calling this section Stronger Together. We're going to examine… that really should be a question mark there, not a colon after that together, because… It's a thesis. We're going to see if we believe it. We're going to discuss it, and just see where our minds take us today. This is a session from the ADAPT Task Force, where we're looking at issues that were originally started bubbling up to us as use cases from the Cognitive and Learning Accessibility Task Force, also an APA task force, and we're trying to solve some of those problems in systematic ways, and we think that there is a nexus, a symbiosis opportunity between what is known as agentic AI and something that we've developed in ADAPT that we're calling Discoverable Destinations. We're going to define both of those for you a little bit in this hour, and we're hopefully mainly gonna talk To one another and discuss what we think about these opportunities. So, with that, as a brief introduction, let me ask the facilitator of our ADAPT Task Force, Lionel Wolberger, to take us through the next steps of where we're going today. And following Lionel, Abhinav will look at some of the use cases that we think are, in fact, a good symbiotic opportunity between agentic AI and discoverable destinations. Lyle, over to you.

Lionel Wolberger: Thank you very much, Janina. Abhinav, next slide. We have a agenda. We're really looking for discussion, but we felt that we needed to lay out the issue. So I'm going to talk about our scope, how we look at Agentic AI, and we'll give an example. We'll talk about how we understand current agents are working, and then we'll get to the point, which is discoverable destinations. We will introduce you to it, explain it, and show how we feel it overlaps with agents, enough that we wanted a broader discussion with the community. So, if you can go to the next slide…

Lionel Wolberger: Agentic AI is a hot topic of today. I understand it's driving the global economy. We have a critical need for standardizing communication protocols and security mechanisms around this vast influx of robots prowling the web. There are all kinds of protocols that we've read about, and people working on them are focusing on interoperability, scalability, collective intelligence.

Lionel Wolberger: We on the web, of course, are interested in security and data leakage, etc. And why does all this come up? Because the autonomous systems that plan, reason, and execute tasks, they want to operate on minimal user input and automate multi-step workflows. And they understand natural language and act on behalf of users. That's how we defined how we see Agentic AI, and next slide.

Lionel Wolberger: There's a general pattern that we've seen. Accessibility features designed for disabled users often create universal benefits, like a curb cut that helps everyone from wheelchair users to parents with strollers. Clear semantics, structured navigation, and machine-readable cues improve the web for assistive technologies and simultaneously empower agentic AI to interpret and act reliably. By making digital environments more predictable and inclusive, these enhancements reduce friction for all users. And can make smarter, safer automation. So, to make things clear, we'd like to turn to an example, and we thought of a very simple example, an agentic AI trying to go out on the web and look at Excel accessibility statements. Notice we've left out of here any value statement, but there could be value to this, like somebody might be looking for a company.

Room 302: That's more accessible than another company.

Lionel Wolberger: Or a host hotel that's more accessible. There could be all kinds of motivations, but notice we've left that out. We've just reduced it to a simple task. Compare accessibility statements across 50 partner websites. It's my pleasure to introduce my colleague on the task force, Abhinav, who works at SAP, and he works on agentic AI in his day job. And he'll be now explaining what a flow would look like for such an agent.

Abhinav Kumar: Thank you, Lionel. So, this is Abhinav, I work for SAP Labs in India. I hope you're able to hear me clearly.

Lionel Wolberger: I hear you well.

Abhinav Kumar: Perfect. Building upon the example that Lionel introduced, there is a user who wants to compare the accessibility statement across 50 partner websites. If you see, as a manual step, the user has to go to the 50 websites, then he or she has to find out the page where this accessibility statement is written, then he has to take all of it, and then he has to create a matrix comparing in those accessibility statements. But let's take that this job is given to our agent, then how agent will work? It will mimic the user behavior. It will go and find out the relevant pages, so basically, in this case, the relevant accessibility statement pages. Post that, it will navigate to those pages, and then it will extract the content and summarize the accessibility statement or the page content found using the LLM's foundational model.

Abhinav Kumar: Now moving ahead, how does it work for the agent currently, at least for the navigation part? There are basically two paradigms. In the one paradigm, the agent has basically the well-defined APIs, and then it can use those APIs to figure out the pages which are available, then it can use the APIs to navigate to those pages and fetch the content. That's very good, but most of the websites won't have these kind of APIs available. In such cases, this is, basically, the agent has to go for a heuristic method. Basically, they need to do either HTML parsing or CSS parsing to find out the possible pages which might be pointing to this accessibility. So maybe it will look for the content where the accessibility statement or similar terms are used, and then it will find where these is page as a site. The other option is if these 50 websites are known, then there can be a site-specific scripts written specifically for these sites, so that the agent can work reliably. But this has a very specific problem, that these are very brittle implementation, and if there is a design change, then everything can stop working. The agents which were working probably till yesterday with a new website design update, they will start filling up. And make the agents work reliably is a very highly unreliable and high-maintenance job. And the crux of the problem is there is no universal navigation standard available.

Abhinav Kumar: So, ADAPT has been working upon a concept called as Discoverable Destinations which will provide the semantic markers for the common pages. Now, I hand it over back to Lionel to explain about the discoverable destinations, and how we think that this discoverable destination can help the agents to solve this problem.

Lionel Wolberger: Yes, thank you, Abhinav. As Janina pointed out in the introduction, the ADAPT task force has taken up accessibility goals that have been in motion for decades, and many of them were first curated by the, COGA, Cognitive and Learning Disabilities task force. One of those was to provide a simpler way to discover well-known, or expected, or generally used destinations. For example, if you go to an e-commerce site, There's a very cluttered page, generally. They throw everything at the end user, because the… Most often thinking of a sighted user, and they want to throw visual cues, and put up the latest sale, and all kinds of things.

Lionel Wolberger: For somebody with cognitive disabilities or neurodiversity, it can be very confusing and difficult to find the login page, the search, the contact us, the store hours. These are all generally expected, affordances or pages that, due to the age of the internet and societal expectations, we expect now to find pages like that, and you will find them, generally. Most websites will have them. Discoverable destinations is a standardized semantic marker for these common pages. For example, a help page, contact, login, accessibility statement, and it would enable predictable navigation. Next slide.

Lionel Wolberger: We turn now to the topic of our discussion. How are we stronger together, agents and discoverable destinations? Let's fast forward a year, let's say our ADAPT task force has succeeded, and websites are enabling the discoverable destinations procedure, which, by the way, it's not a new standard, it's just a way that a website can put its destinations semantically identified. Kind of like a robots.txt will show pages that robots can scan. This is a list of expected destinations. So, you would discover the destinations via semantic identifiers. What would work for a disabled person would also work for an agentic robot. It could navigate reliably to those key pages, and this would enable consistent automation across compliant sites. Next slide.

Lionel Wolberger: We're rounding out to the end of our introduction and looking forward to a conversation over this. So, we thought ahead, the reason why we asked for the session is, how could LLMs use such a markup? Well, the LLM is doing a planning, reasoning, and content synthesis. They would rely on semantic tools based on our discoverable destinations procedure to discover the pages, to navigate to them, and retrieve content. And then, because these are pages that invite action, they could take actions. They could do password changes, they could combine destinations and APIs. So I'll go back to Abhinav.

Lionel Wolberger: To finish up our introduction, and then we'll be turning to discussion. Ahbinav?

Abhinav Kumar: Yeah, thanks, Lionel. So, just to add upon it, so basically, this workflow, what we are talking about, solves the problem of navigation and content retrieval. But in case of complex things, like the password change, probably the user might require to provide its previous password, or it might be requiring a two-factor authentication. In those cases, basically, it will be a combination of these destinations, plus probably human in-loop or APIs, where the destination will take you to the pages, like, for the password change, and then the user can provide its pass, password, and in case of two-factor authentication and other things. Basically, in the complex cases, basically humans plus these destinations will work together.

Abhinav Kumar: Moving ahead, for… with the sample workflow, for the example which we took, that we are trying to compare the accessibility statement across 50 partner websites. So, how it will start, that the agent will try to retrieve the list of the discoverable destinations provided by these 50 websites. Then it will search for the relevant destination, which in our case is basically the accessibility statement. And if found, it will navigate to those destinations, and if not found, then basically to notify the user, and then stop.

Abhinav Kumar: Pardon? I heard some noise. Is there a question?

Lionel Wolberger: No, you can go ahead.

Room 302: Okay. Incidental. Incidental, sorry.

Abhinav Kumar: And in case, if the destination is there, it can either navigate, and if the task requires to fetch the content, like in our case, to compare the accessibility statement, then the content is retrieved, and then it is processed and summarized using the foundational model and presented to the user.

Abhinav Kumar: With this, we have come to the end of our presentation, and basically, we want to discuss with you, as we already talked in the starting, that how do you see whether this adapt at discoverable destinations, and our agent API coming together for a better experience for the agents.