10:18:12 RRSAgent has joined #ai-agents 10:18:16 logging to https://www.w3.org/2025/03/26-ai-agents-irc 10:18:16 RRSAgent, do not leave 10:18:17 RRSAgent, this meeting spans midnight 10:18:17 RRSAgent, make logs public 10:18:18 Meeting: How would AI Agents change the Web platform? 10:18:18 Chair: Dominique Hazaël-Massieux 10:18:18 Agenda: https://github.com/w3c/breakouts-day-2025/issues/7 10:18:18 Zakim has joined #ai-agents 10:18:19 Zakim, clear agenda 10:18:19 agenda cleared 10:18:19 Zakim, agenda+ Pick a scribe 10:18:21 agendum 1 added 10:18:21 Zakim, agenda+ Reminders: code of conduct, health policies, recorded session policy 10:18:21 agendum 2 added 10:18:21 Zakim, agenda+ Goal of this session 10:18:22 agendum 3 added 10:18:22 Zakim, agenda+ Discussion 10:18:22 agendum 4 added 10:18:22 Zakim, agenda+ Next steps / where discussion continues 10:18:24 agendum 5 added 10:18:25 Zakim, agenda+ Adjourn / Use IRC command: Zakim, end meeting 10:18:25 agendum 6 added 10:18:25 breakout-bot has left #ai-agents 10:26:20 tidoust has joined #ai-agents 13:21:04 tantek has joined #ai-agents 13:40:13 ktoumura has joined #ai-agents 13:53:10 ktoumura has joined #ai-agents 13:53:10 tantek has joined #ai-agents 13:59:24 bkardell_ has joined #ai-agents 14:00:02 hi! 14:00:04 Present+ Brian, Anssi, Hadley, DavidBaron, PLH, Dom 14:00:56 dbaron has joined #ai-agents 14:01:03 danbri has joined #ai-agents 14:01:15 hadleybeeman has joined #ai-agents 14:01:15 denis has joined #ai-agents 14:01:15 plh has joined #ai-agents 14:01:20 present+ 14:01:37 zakim, who is here? 14:01:37 Present: Brian, Anssi, Hadley, DavidBaron, PLH, Dom 14:01:38 On IRC I see plh, denis, hadleybeeman, danbri, dbaron, bkardell_, tantek, ktoumura, tidoust, Zakim, RRSAgent, dom 14:02:04 hyojin has joined #ai-agents 14:02:21 present+ hadleybeeman , dbaron , ivan_herman 14:02:24 anssik has joined #ai-agents 14:02:36 present+ Kunihiko_Toumura 14:02:37 Present+ DanBri 14:03:03 Present+ Anssi_Kostiainen 14:03:05 Present+ Hyojin Song 14:03:10 Mizushima has joined #ai-agents 14:03:23 present+ Larry_Zhao 14:03:43 present+ 14:03:43 present+ Iris_ben 14:03:47 present+ Scott_Hendrickson 14:03:53 present+ Ari 14:04:07 present+ Kaustubha 14:04:08 scribe+ 14:04:16 McCool has joined #ai-agents 14:04:32 present+ Tomoaki_Mizushima 14:04:33 passcode in invite seems to be incorrect for zoom 14:04:44 ivan has joined #ai-agents 14:05:18 -> https://www.w3.org/2025/Talks/dhm-ai-agents/ Slides 14:05:27 [slide 2] 14:05:42 Dom: AI is in the news noawadays 14:05:48 RobSmith has joined #ai-agents 14:05:51 ... agentic AI, etc. 14:05:54 passcode in invite is incorrect 14:06:22 passcode 075582 14:06:41 thx 14:06:47 [slide 3] 14:07:08 dom: let's a look at how LLMs interact on the Web 14:07:10 [slide 4] 14:07:32 dom: chatbot is simple and classic 14:07:36 [slide 5] 14:07:50 dom: evolution: LLM on the client side 14:08:04 ... the web site download an LLM in your browser 14:08:11 ... (WebGPU, WebNN, WASM) 14:08:25 [slide 6] 14:08:56 dom: variation: Task APIs. You interact with the embedded LLM, which you access through well scoped APIs 14:09:16 ... the LLM is operating on the user device and linked to specific tasks 14:09:20 [slide 7] 14:09:38 dom: you delegate to the LLM 14:10:00 ... instead of searching for a shop to buy shoes 14:10:16 ... you ask the LLM to do it for you 14:10:47 ... so analizing search results, initiating the purchase, etc. 14:11:20 ... so the browser is no longer the core center of user interaction 14:11:30 ... the LLM uses the browser 14:11:36 [slide 8] 14:12:02 dom: the LLM could on server-side for those cases 14:12:07 s/coul/could be/ 14:12:30 ... so those were appromixation of LLM variations 14:12:36 [slide 9] 14:12:51 dom: home assistants have been deployed for sometime 14:13:06 ... those had some charateristics of AI agents 14:13:21 ... some are shipping chatbots as part of the UI 14:13:33 ... interacting with the webpage as part of the browser 14:13:33 (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 14:13:33 channel binding. 14:13:54 ... recent announcement from firefox having an LLM in a browser extension 14:14:33 ... [deep research agents] 14:15:04 ... uses AI agent to test web content 14:15:17 q+ 14:15:41 ack mc 14:16:04 McCool: an other option on architecture: using a nearby computer 14:16:19 ... has some privacy/security advantages 14:16:49 handy9 has joined #ai-agents 14:16:50 ... giving raw access to AI agents to your browsing has privacy implications 14:17:13 [slide 10] 14:17:39 dom: create additional intermediate between the user and the web 14:17:58 We the TAG have been talking a lot about when this becomes a user agent. 14:18:17 ... switching from clicking to access services to a descriptive interaction model 14:18:42 ... non-deterministic: output will differ 14:19:29 ... boundaries between robots and user agents. robots are fully automated, while a browser is controlled by a user 14:19:55 ... AI agent is not quite a crawler but not exactly a user agent either 14:20:02 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) 14:20:05 ... should those agents respect robots.txt ? 14:20:16 [slide 11] 14:20:30 dom: impact on web platform 14:20:44 ... MCP to interact with online services 14:21:09 ... it's an emerging specification 14:21:39 present+ 14:21:40 ... a new of website are looking at providing LLM-tailored content 14:22:06 ... seems similar to early discussions on mobile and a11y 14:22:23 ... efforts looking at sitemaps for AI agents 14:22:25 LLMs.txt spec - https://llmstxt.org/ 14:22:36 ... content gets customized for LLMs 14:23:16 ... LLM and CATPCHA: LLMs may prevented from solving those 14:23:20 cpn has joined #ai-agents 14:23:25 [slide 12] 14:23:34 dom: risks 14:24:06 ... security: hallucinations: LLM may be doing something as completely different than expected. the more access, the more risk... 14:24:48 ... no strong difference between what they're supposed to do as a program vs on behalf of the user 14:25:09 ... eg sending your credentials to third party 14:25:20 ... privacy 14:25:40 ... with distributed architecture, it may make privacy more complex 14:25:59 ... ecosystem 14:26:16 ... diluted the end user intent from the content provider 14:26:38 ... content is monetized based on the assumption that ads will be seen by users 14:26:51 .... if you reduce it, it will change the dynamics 14:27:02 [slide 13] 14:27:16 dom: what did I miss? 14:27:28 ... input on the role of W3C in this space? 14:27:37 ... how to organize a discussion 14:28:01 Dingwei has joined #ai-agents 14:28:03 q+ to mention multimodal models (beyond LLM text only) and webrtc 14:28:03 ... should we wait before it's a fast moving space? 14:28:20 q+ to talk about MCP 14:28:49 ... if you have contacts for us, we're interested 14:29:00 [end of presentation] 14:29:03 q? 14:29:04 q? 14:29:06 q+ 14:29:10 ack dan 14:29:10 danbri, you wanted to mention multimodal models (beyond LLM text only) and webrtc 14:29:21 danbri: nice to see you're on top of all of this. 14:30:05 zolkis has joined #ai-agents 14:30:11 q+ 14:30:15 q+ 14:30:19 ... 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 14:30:26 ... using raw sockets 14:30:47 ... my advise would be to build on top of those uses 14:31:08 ... rather rather the general AI 14:31:17 (use cases would be helpful - e.g. accessibility, describe images; language transcription; PNG->SVG conversion; etc; BUT evolving quickly; also, not just LLMs) 14:31:19 dom: taking into account greater diversity of needs 14:31:20 webrtc openai native support, https://platform.openai.com/docs/guides/realtime - webrtc for realtime api (also websockets); 14:31:20 3p gemini integration https://github.com/pipecat-ai/gemini-webrtc-web-simple https://ai.google.dev/gemini-api/docs/live 14:31:22 scribe+ 14:31:49 PLH: my understanding of MCP is that it is not meant to replace HTTP, but is an RCP protocol to be used locally 14:32:09 ... it comes with no security layer - e.g. it reuses the user credentials 14:32:47 dom: where do you put the boundaries? 14:32:50 q? 14:32:55 ... if browsers have an LLM 14:32:56 scribe- 14:33:01 ack plh 14:33:01 plh, you wanted to talk about MCP 14:33:05 ack anssik 14:33:19 tomayac has joined #ai-agents 14:33:25 anssi: the potential to disrupt the current moneytization of the web 14:33:27 present+ 14:33:39 ... "nobydo knows you're an AI agent" 14:33:59 ... but a lot of positive here, useful use cases. 14:34:21 ... it's still valuable for the community to keep up with this development 14:34:35 .... standardize pieces maybe 14:34:49 ... w3c cares about horizontals 14:35:01 ... ethical being one of them 14:35:09 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. 14:35:09 """Our top priority is improving remote MCP connections, allowing clients to securely connect to MCP servers over the internet. Key initiatives include: 14:35:09 Authentication & Authorization: Adding standardized auth capabilities, particularly focused on OAuth 2.0 support. 14:35:09 Service Discovery: Defining how clients can discover and connect to remote MCP servers. 14:35:10 https://www.gov.uk/government/publications/international-ai-safety-report-2025 14:35:11 Stateless Operations: Thinking about whether MCP could encompass serverless environments too, where they will need to be mostly stateless.""" 14:35:36 McCool: looking at use cases and gaps into our specifications 14:35:59 Harry4 has joined #ai-agents 14:36:09 .... writing use cases and thinking their impact on use cases would be useful. not sure who should do it. 14:36:20 ... lots of positive around a11y 14:36:29 q+ 14:36:38 dom: collecting use cases with a cg? 14:36:46 ... I did not include a11y in my slides 14:36:55 ... but agree it's an important aspect 14:36:56 ack McCool 14:37:09 McCool: image description, reorganizing layouts, etc. 14:37:18 ... but it's opended and it is evolving 14:37:28 ack zolkis 14:37:46 zolkis: user agents vs AI agents, AI workflows, normal workflows 14:38:14 ... in a web browser we need to think about process architecture changes vs thinking about workflows 14:38:28 ... not clear where AI agents fall into those spaces 14:38:54 dom: good input to look into. I welcome pointers 14:39:05 hadleybeeman: [muted] 14:39:12 RRSAgent, draft minutes 14:39:14 I have made the request to generate https://www.w3.org/2025/03/26-ai-agents-minutes.html dom 14:39:21 ... [still muted] 14:39:30 ... [lots of good things that we can't hear] 14:39:43 shendrickson has joined #ai-agents 14:40:19 .... thanks for the work done so far 14:40:31 ... TAG has AI in its radar 14:40:41 ... and what can we say and do 14:40:52 ... we don't have consensus yet on a statement but working towards it 14:41:28 ... I'm focusing otm: LLMs as user agents 14:41:55 .... they should have the same responsibilities than user agents 14:41:55 [related to User Agents duties & TAG: https://w3ctag.github.io/user-agents/ ] 14:41:59 ... going back to ethical web principles and web design principles 14:42:21 ... making sure those can adapt to new technologies 14:42:29 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 14:42:43 ... wanted to hear what I'm missing and what else others are doing 14:42:49 q+ 14:43:03 dom: where should we continue the conversation? 14:43:10 McCool: should we have a workshop? 14:43:16 q+ 14:43:34 dom: i've been struggling to find the right way to move forward, rather than just watching 14:43:43 ... re workshop: not clear to me that we know enough to have a productive conversation 14:43:49 q+ 14:43:56 McCool: one breakout isn't enough 14:44:05 .... workshop, cg, tpac 14:44:14 ack hadleybeeman 14:44:16 ack McCool 14:44:25 ack anssik 14:44:30 anssik: I can talk to the web machine learning wg/cg 14:44:35 .... to see if there is interest 14:44:48 ... also we can use time in the cg if you would like 14:44:54 I should’ve mentioned the Web ML group — also written wonderful docs. I got a lot out of the use cases. 14:45:07 ... interested in hearing positions from the browser vendors or other shipping products 14:45:22 q+ to mention AC meeting 14:45:29 q+ to suggest a WG-centric workflow for a W3C response to the AI situation 14:46:24 shendrickson: on progressing this work: IETF looking into authenticating agents to web sites 14:46:39 .... we should try to find a few more of those 14:46:46 ... but space is very broad 14:47:01 .... we should try to focus on use cases around agents 14:47:25 dom: intersection is indeed difficult 14:47:25 ack shendrickson 14:47:27 ack plh 14:47:27 plh, you wanted to mention AC meeting 14:47:33 scribe+ 14:47:49 PLH: I want an AI agent to scribe in my stead 14:47:59 ... there will be another AI-related session at the upcoming AC meeting 14:48:17 scribe- 14:48:20 ack danbri 14:48:20 danbri, you wanted to suggest a WG-centric workflow for a W3C response to the AI situation 14:48:37 danbri: this trend is as big as when the web started. all groups should think about it 14:48:51 ... I would encourage to have a top-level to grassroot approach 14:48:52 q+ to mention parallels with search engines 14:49:13 ... asking every group for a one-page on how AI impacts them 14:49:34 dom: I love the suggestion. it grounds to our current work 14:49:56 ... for AI in general or for AI agent evolution? 14:50:09 danbri: for everything 14:50:21 ... is there anyone thinking about forms? 14:50:41 dom: this might also serve as a good ground before a workshop 14:51:07 RobSmith: parallels with existing systems: sitemaps, etc. 14:51:21 ... search engines are another one 14:51:29 ... SEOs optimization 14:51:39 ... taking some parallel approachs for AI 14:51:57 ... AI is an extension of robots 14:52:19 (also - semantic tagging...) 14:52:27 dom: comparison with search engines is listed in the AI impact report 14:52:43 ... whether robots.txt should be used to control AI robots 14:52:53 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 14:53:12 .... but search engines operated on prefetch content, while AIs are doing live-fetch content 14:53:39 ... open sources complaining about AI agents fetching documentation, etc. 14:54:02 q? 14:54:08 ... the more we can anchor ourselves in problems we already know, the better 14:54:15 ack RobSmith 14:54:15 RobSmith, you wanted to mention parallels with search engines 14:54:31 RobSmith: correct about pfretech/live, avoiding DOS attacks 14:54:45 dom: it's distributed DOS :( 14:54:57 dom: in conclusion 14:55:04 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 14:55:04 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." 14:55:04 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!) 14:55:07 ... upcoming AC meeting has a session on AI 14:55:31 ... will look into danbri suggestion on building a broader view from all of our groups to AI impact 14:55:45 ... not ready to commit to a workshop or a cg, but get in touch with me 14:55:54 ... very interested in contacts 14:56:08 ... dom@w3.org 14:56:19 ... anything else? 14:56:20 q+ 14:56:54 anssi: should we aim to create a web platform where you can't distinguish between humans and AIs? 14:57:00 Can we stop it happening? 14:57:19 dom: also, should we be able to detect it? 14:57:32 ... very interesting way of framing the debate 14:57:52 danbri: get and post are very different 14:58:05 ... more interesting in dinstinguishing the post than the get 14:58:15 dom: feel free to get in touch with me 14:58:42 [end of discussion] 14:58:47 ivan has left #ai-agents 14:59:01 RRSAgent, draft minutes 14:59:02 I have made the request to generate https://www.w3.org/2025/03/26-ai-agents-minutes.html dom 14:59:46 i|-> https://www.w3.org/2025/Talks/dhm-ai-agents/ Slides|Slideset: https://www.w3.org/2025/Talks/dhm-ai-agents/ 14:59:48 RRSAgent, draft minutes 14:59:50 I have made the request to generate https://www.w3.org/2025/03/26-ai-agents-minutes.html dom 15:01:36 denis has left #ai-agents 15:02:43 hyojin has left #ai-agents 15:21:15 arichiv has joined #ai-agents 22:58:01 RRSAgent, bye 22:58:01 I see no action items