W3C

– DRAFT –
WAI-Adapt Teleconference

26 Aug 2025

Attendees

Present
Abhinav, janina, Lionel_Wolberger, matatk
Regrets
Annalu
Chair
Lionel_Wolberger
Scribe
Lionel_Wolberger, janina, matatk

Meeting minutes

Symbols

matatk: I am ready to move the Symbols explainer forward this week

Lionel_Wolberger: We'll try and get you some time back for that

Destinations

<matatk> General progress update: https://lists.w3.org/Archives/Public/public-adapt/2025Aug/0004.html - I didn't mention in here that I also updated the extension to Google's new Manifest V3 API, as well as bumping the packages.

Lionel_Wolberger: Is look and feel the same?

matatk: May go with sidebar rather than popup.

Well-known destinations and machine-readable semantics

<matatk> rrsagnet, make minutes

Abhinav: I am considering how agentic AI would interpret our destinations

<matatk> https://en.wikipedia.org/wiki/Agentic_AI

Abhinav: in my experience with recent LLM models and their reasoning
… take for example a flight booking: I have some conditions, an overall travel time and a date range
… an LLM can parse that out into source, destination, date range and constraints
… it likely has a toolkit that it then draws upon, for example an API to a flight information service
… and then another tool that takes that data, compares and filters as to the conditions it received
… the filtering may be iterative: one filter for price range, one for flight duration
… followed by another layer that may be able to use my credit card to book it

<matatk> janina: *mentions https://innosearch.ai *

janina: Innosearch is an example of such a service, it uses expedia on its backend

matatk: While we will likely implement semantics to aid in machine readability
… but I note that visual LLMs are available to do the job without semantics - i.e. more like browsing sites visually

Lionel_Wolberger: These visual presentations are coherent, typically thanks to underlying semantics

<matatk> matatk: MCP that Abhinav mentioned: https://modelcontextprotocol.io/docs/getting-started/intro

Abhinav: Model context protocol was put forward by Anthropic, and is commonly used by these types of agents
… it is a method to expose your LLM to such external services
… it enables an Agent to select the proper tool. The MCP exposes the tools to the Agents
… the previous generation of web agents, that use APIs to a website, can also use this MCP
… and these web agents can be called via MCP
… the MCP requires authorization
… Our approach to destinations is very much required in this context.
… Consider (a) AI Agents (b) tools (c) MCP interface
… these agents are struggling to find the relevant pages, via DOM parsing or visual inspection to find the navigation
… In my example, if a WKD clarifies that a website has an endpoint for flight search, the agent will go to that endpoint to get its results

janina: We should chase this at TPAC.

Abhinav: In summary, this WKD concept is supporting the current transition to Generative AI optimization

janina: Because we are making the URLs declarative

janina: Accessibility is more and more important, as accessibility relies on semantics and regularities that also benefit the LLMs

matatk: The 'destination' is the linkset
… checkout is more of an action than a destination
… it's a thing that you do, rather than a place that you go
… Abhinav explained this as dependent on whether you have something
… so we may need to revisit action v destination
… also, janina brought up payments
… there's payment for the final item, what you are buying; there's also paying for API usage
… web payments is for the former not the latter

Lionel_Wolberger: I like it that 'actions' would be useful for LLMs. that could motivate us to tackle that next

Abhinav: Should we add this LLM agentic consideration to the explainer?

matatk: +1

Lionel_Wolberger: +1

matatk: A short addition to the use cases section is a good place to start

Minutes manually created (not a transcript), formatted by scribe.perl version 244 (Thu Feb 27 01:23:09 2025 UTC).

Diagnostics

Succeeded: s/seem to be doing a reasonable/are available to do the/

Succeeded: s/job without semantics/job without semantics - i.e. more like browsing sites visually/

All speakers: Abhinav, janina, Lionel_Wolberger, matatk

Active on IRC: Abhinav, janina, Lionel_Wolberger, matatk