– DRAFT –
How would AI Agents change the Web platform?

<bkardell_> hi!

<McCool> passcode in invite seems to be incorrect for zoom

Slideset: https://www.w3.org/2025/Talks/dhm-ai-agents/

Slides

Dom: AI is in the news noawadays
… agentic AI, etc.

<McCool> passcode in invite is incorrect

passcode 075582

<McCool> thx

dom: let's a look at how LLMs interact on the Web

dom: chatbot is simple and classic

dom: evolution: LLM on the client side
… the web site download an LLM in your browser
… (WebGPU, WebNN, WASM)

dom: variation: Task APIs. You interact with the embedded LLM, which you access through well scoped APIs
… the LLM is operating on the user device and linked to specific tasks

dom: you delegate to the LLM
… instead of searching for a shop to buy shoes
… you ask the LLM to do it for you
… so analizing search results, initiating the purchase, etc.
… so the browser is no longer the core center of user interaction
… the LLM uses the browser

dom: the LLM could bed on server-side for those cases
… so those were appromixation of LLM variations

dom: home assistants have been deployed for sometime
… those had some charateristics of AI agents
… some are shipping chatbots as part of the UI
… interacting with the webpage as part of the browser

<danbri> (minor nitpic) re MCP, it's not specifically an http API, rather they initially define only stdio and HTTP transports https://modelcontextprotocol.io/docs/concepts/architecture ... initially same local host, with a view to extending to networked services later. It's basically some JSON-RPC interfaces. I've been experimenting with a WebRTC data

<danbri> channel binding.

dom: recent announcement from firefox having an LLM in a browser extension
… [deep research agents]
… uses AI agent to test web content

McCool: an other option on architecture: using a nearby computer
… has some privacy/security advantages
… giving raw access to AI agents to your browsing has privacy implications

dom: create additional intermediate between the user and the web

<hadleybeeman> We the TAG have been talking a lot about when this becomes a user agent.

dom: switching from clicking to access services to a descriptive interaction model
… non-deterministic: output will differ
… boundaries between robots and user agents. robots are fully automated, while a browser is controlled by a user
… AI agent is not quite a crawler but not exactly a user agent either

<McCool> two main points: where the LLM runs, and what it has access to (which has two aspects: what data it can use, and what it can do - delegated rights, etc)

dom: should those agents respect robots.txt ?

dom: impact on web platform
… MCP to interact with online services
… it's an emerging specification
… a new of website are looking at providing LLM-tailored content
… seems similar to early discussions on mobile and a11y
… efforts looking at sitemaps for AI agents

<danbri> LLMs.txt spec - https://llmstxt.org/

dom: content gets customized for LLMs
… LLM and CATPCHA: LLMs may prevented from solving those

dom: risks
… security: hallucinations: LLM may be doing something as completely different than expected. the more access, the more risk...
… no strong difference between what they're supposed to do as a program vs on behalf of the user
… eg sending your credentials to third party
… privacy
… with distributed architecture, it may make privacy more complex
… ecosystem
… diluted the end user intent from the content provider
… content is monetized based on the assumption that ads will be seen by users
… if you reduce it, it will change the dynamics

dom: what did I miss?
… input on the role of W3C in this space?
… how to organize a discussion
… should we wait before it's a fast moving space?
… if you have contacts for us, we're interested

[end of presentation]

<Zakim> danbri, you wanted to mention multimodal models (beyond LLM text only) and webrtc

danbri: nice to see you're on top of all of this.
… it's good to focus on multimodal. some of those can understand images, sounds, etc. streaming voice system from google, etc. they're starting to use webrtc
… using raw sockets
… my advise would be to build on top of those uses
… rather rather the general AI

<McCool> (use cases would be helpful - e.g. accessibility, describe images; language transcription; PNG->SVG conversion; etc; BUT evolving quickly; also, not just LLMs)

dom: taking into account greater diversity of needs

<danbri> webrtc openai native support, https://platform.openai.com/docs/guides/realtime - webrtc for realtime api (also websockets);

<danbri> 3p gemini integration pipecat-ai/gemini-webrtc-web-simple https://ai.google.dev/gemini-api/docs/live

PLH: my understanding of MCP is that it is not meant to replace HTTP, but is an RCP protocol to be used locally
… it comes with no security layer - e.g. it reuses the user credentials

dom: where do you put the boundaries?
… if browsers have an LLM

<Zakim> plh, you wanted to talk about MCP

anssi: the potential to disrupt the current moneytization of the web
… "nobydo knows you're an AI agent"
… but a lot of positive here, useful use cases.
… it's still valuable for the community to keep up with this development
… standardize pieces maybe
… w3c cares about horizontals
… ethical being one of them

<danbri> on plh's MCP summary, it captures where MCP is today but there's a roadmap https://modelcontextprotocol.io/development/roadmap and non-local non-public services are clearly on their way.

<danbri> """Our top priority is improving remote MCP connections, allowing clients to securely connect to MCP servers over the internet. Key initiatives include:

<danbri> Authentication & Authorization: Adding standardized auth capabilities, particularly focused on OAuth 2.0 support.

<danbri> Service Discovery: Defining how clients can discover and connect to remote MCP servers.

<anssik> https://www.gov.uk/government/publications/international-ai-safety-report-2025

<danbri> Stateless Operations: Thinking about whether MCP could encompass serverless environments too, where they will need to be mostly stateless."""

McCool: looking at use cases and gaps into our specifications
… writing use cases and thinking their impact on use cases would be useful. not sure who should do it.
… lots of positive around a11y

dom: collecting use cases with a cg?
… I did not include a11y in my slides
… but agree it's an important aspect

McCool: image description, reorganizing layouts, etc.
… but it's opended and it is evolving

zolkis: user agents vs AI agents, AI workflows, normal workflows
… in a web browser we need to think about process architecture changes vs thinking about workflows
… not clear where AI agents fall into those spaces

dom: good input to look into. I welcome pointers

hadleybeeman: [muted]
… [still muted]
… [lots of good things that we can't hear]
… thanks for the work done so far
… TAG has AI in its radar
… and what can we say and do
… we don't have consensus yet on a statement but working towards it
… I'm focusing otm: LLMs as user agents
… they should have the same responsibilities than user agents

<dom> [related to User Agents duties & TAG: https://w3ctag.github.io/user-agents/ ]

hadleybeeman: going back to ethical web principles and web design principles
… making sure those can adapt to new technologies

<zolkis> A good primer/summary (further material on the channel) on AI workflows vs AI agents (and you can ask Gemini/Perplexity/ChatGPT as well): https://www.youtube.com/watch?v=tx5OapbK-8A

hadleybeeman: wanted to hear what I'm missing and what else others are doing

dom: where should we continue the conversation?

McCool: should we have a workshop?

dom: i've been struggling to find the right way to move forward, rather than just watching
… re workshop: not clear to me that we know enough to have a productive conversation

McCool: one breakout isn't enough
… workshop, cg, tpac

anssik: I can talk to the web machine learning wg/cg
… to see if there is interest
… also we can use time in the cg if you would like

<hadleybeeman> I should’ve mentioned the Web ML group — also written wonderful docs. I got a lot out of the use cases.

anssik: interested in hearing positions from the browser vendors or other shipping products

shendrickson: on progressing this work: IETF looking into authenticating agents to web sites
… we should try to find a few more of those
… but space is very broad
… we should try to focus on use cases around agents

dom: intersection is indeed difficult

<Zakim> plh, you wanted to mention AC meeting

PLH: I want an AI agent to scribe in my stead
… there will be another AI-related session at the upcoming AC meeting

<Zakim> danbri, you wanted to suggest a WG-centric workflow for a W3C response to the AI situation

danbri: this trend is as big as when the web started. all groups should think about it
… I would encourage to have a top-level to grassroot approach
… asking every group for a one-page on how AI impacts them

dom: I love the suggestion. it grounds to our current work
… for AI in general or for AI agent evolution?

danbri: for everything
… is there anyone thinking about forms?

dom: this might also serve as a good ground before a workshop

RobSmith: parallels with existing systems: sitemaps, etc.
… search engines are another one
… SEOs optimization
… taking some parallel approachs for AI
… AI is an extension of robots

<McCool> (also - semantic tagging...)

dom: comparison with search engines is listed in the AI impact report
… whether robots.txt should be used to control AI robots

<danbri> at some point Google started using a new string for AI usecases, https://developers.google.com/search/docs/crawling-indexing/overview-google-crawlers ... not sure if that got rolled back

dom: but search engines operated on prefetch content, while AIs are doing live-fetch content
… open sources complaining about AI agents fetching documentation, etc.
… the more we can anchor ourselves in problems we already know, the better

<Zakim> RobSmith, you wanted to mention parallels with search engines

RobSmith: correct about pfretech/live, avoiding DOS attacks

dom: it's distributed DOS :(

dom: in conclusion

<danbri> see https://developers.google.com/search/docs/crawling-indexing/google-common-crawlers "Google-Extended is a standalone product token that web publishers can use to manage whether their sites help improve Gemini Apps and Vertex AI generative APIs, including future generations of models that power those products. Grounding with Google Search on

<danbri> Vertex AI does not use web pages for grounding that have disallowed Google-Extended. Google-Extended does not impact a site's inclusion or ranking in Google Search."

<danbri> This is for sites to use robots.txt to exclude crawlers that feed certain AI apps without dropping out of general search (afaik; I am no longer at GOogle and do not speak for them!)

dom: upcoming AC meeting has a session on AI
… will look into danbri suggestion on building a broader view from all of our groups to AI impact
… not ready to commit to a workshop or a cg, but get in touch with me
… very interested in contacts
… dom@w3.org
… anything else?

anssi: should we aim to create a web platform where you can't distinguish between humans and AIs?

<hadleybeeman> Can we stop it happening?

dom: also, should we be able to detect it?
… very interesting way of framing the debate

danbri: get and post are very different
… more interesting in dinstinguishing the post than the get

dom: feel free to get in touch with me

[end of discussion]