Meeting minutes
Anssi: as a reminder, we'll use IRC-based queue management in this meeting:
https://
Anssi: to suggest agenda topics, use Agenda+ label, e.g.:
Anssi: welcome back everyone, we made significant progress at TPAC and have important follow-ups to discuss today
Anssi: first, please welcome our latest new participants:
… Christopher Wood and Thibault Meunier from Cloudflare
… Artur Serra from Snyk
… Taylor Dolezal from Dosu
… Gary Klassen from Google
… Liad Yosef from Ora
… Peter Rosemann from Alpin Insight Solutions
… Prabhu Rangaraju from Penske Logistics
… Marinoni Moretto, Song Zhang, Daniel Mantei, Steven Burke, Chris McKenzie, Ragdan Abdulbaqi, Sai Vineela Ganti, Ishan Anand, Damisile Ayoola, Dongxun Jin, Juliano Sig, Caio Gondim, Damisile Ayoola, Dongxun Jin, Caio Gondim as individual contributors
… welcome all on board this Community Group
Beyond WebMCP
Anssi: this is a follow up to the TPAC Agentic Web breakout session where we discussed the steps beyond WebMCP:
Anssi: the goal of today's discussion is to advance the following proposed Agentic Web WG's tentative deliverables toward an explainer form:
… - Agent Identify
… - Agent Delegation
… - Agent Interaction Measurement
… the "Agentic Interaction Modes" deliverable needs some clarity wrt normative vs non-normative aspects
Anssi: for context, about Working Group vs Community Group interplay
… when W3C charters a Working Group, the established pattern is to use a Community Group to handle incubation for topics related to the Working Group
… this WebML CG serves that incubation function for WebMCP and related topics
… W3C WG charters are expected to be well-defined wrt normative deliverable scope to allow W3C members evaluate IP commitments
… to that end, this Community Group will produce explainers for the 3 proposed tentative deliverables (Agent Identify, Agent Delegation, Agent Interaction Measurement) to allow the Agentic Web WG charter to make timely progress to the Advisory Committee review
… the explainers and draft specs developed in this Community Group are expected to advance to an appropriate WG after successful incubation
… in a nutshell, the W3C Working Groups are governed by the W3C Process that applies the RF IPR licensing to the resulting specs, the end product aka W3C Recommendation
… future proposals will advance through a Community Group incubation process similarly to WebMCP before moving to a WG for standardization and eventual W3C Recommendation
… to make this work, seamless collaboration between this CG and the proposed Agentic Web WG is required to ensure a smooth transition
Anssi: the explainers should follow the usual structure, concise overview of the problem, proposed solution(s), use cases, alternatives, any known considerations wrt a11y, i18n, privacy, security
"Explainer explainer" document
Anssi: with that backdrop, let's discuss the proposed tentative deliverables, identify owners who take responsibility for transforming the issues into explainers to allow the Agentic Web WG charter advance in a timely manner
… I'll create repos for each proposed tentative deliverable to facilitate focused development and collaboration
… any questions about this general goal?
Mark: is the goal to have explainers for these topics to have them completed by the end of the month?
Anssi: yes
Agent Identity
Anssi: issues webmachinelearning/
<gb> Issue 54 not found
<gb> Issue 365 not found
Anssi: Agent Identity covers a mechanism for sites and agents to convey and negotiate identity with specific attribution of agent-originated interaction, for example:
… "I'm an agent working on a user's behalf"
… "This Submit event was initiated by an agent acting on behalf of a user"
… one proposed solution is to extend the Event interface with a new agentInvoked attribute, good starting point for the explainer
https://
Anssi: Ilya explained the use cases and requirements for the proposed mechanism from Shopify's perspective drawing from large scale deployments, this is additional good input for the explainer document
<gb> Issue 576 [wg/agenticweb] Agentic Web Group Charter (by plehegar) [AI] [Horizontal review requested] [Security review completed] [In Charter Refinement] [charter]
Yoav: essentially I think it would be good to start with an explainer that outlines the problem, then try to map the solutions to the use cases and see what other groups are doing in e.g. IETF
<Mark_Foltz> +1, we need to focus explainers on problem definitions and not solutions in the short time frame
Dom: I agree, we need to understand the problem we want to solve, and when the agent would expose its identity, e.g. through clicks
… I think we want a clearer picture what interface the agent will use to interact with the browser
… we want to figure out the problem we want to solve and interface by which the agent interfaces with the browser
Benjamin: I wanted to take a step back and ask, to bring them in as WG items we need to recharter, I'm questioning why to rush this work, perhaps more natural incubation without rush is wiser
Mark: I raised this as an issue against the draft charter, I wanted to ask what exactly is an agent and what technologies are in scope
Yoav: I agree with Ben's comment
… I do think it is better to tackle these sooner than later, we don't want to rush to solutions, we need a clearer sense of problem statements
Ehsan: I agree with Ben, Yoav and Mark, this discussion needs a lot of pre-discussion before getting to solution, I'm fine to have a document that defines the problem, also parallel discussion expected in WebAuth and elsewhere
<domfarolino> happy to help with this
Repository: webmachinelearning/webmcp
<Idan> I can help with mapping use cases for the explainer
<domfarolino> +1
<yoav> +1
RESOLUTION: Yoav and Dom to draft an explainer for Agent Identity, focusing on problem statement and use cases. (issues webmachinelearning/
Agent Delegation
Anssi: issues #44 and #54
… mechanisms for users to grant agents authority to act on their behalf on a site to take actions
… relevant in this context is to understand implications for different interaction modes
… including enterprise use cases and requirements
… two facets of agent delegation: action-specific permissions and agents as separately authorized clients
… 1) action-specific permissions, issue #44
… "I trust agent 'Foo' to take this action 'Bar' in the future on this site (example.com), don't ask me again"
Khushal: both agent and the site want to ensure they have explicit user's consent, avoid duplicated consenting once for browser, once for agent
Umar: a complicated topic, for AI agents, one of the ideas is to propagate contextual information
… a lot of work on checking user's prior interaction, are they confident making a new decision based on their history
… or what not giving a permission implies
… 2) agents as separately authorized clients, issue #54
<gb> Issue 54 Challenging assumptions of Identity within WebMCP (by EmLauber) [beyond-scope: agents-on-the-web]
Anssi: "agent 'Foo' is authenticated and authorized to call WebMCP tools on this site (example.com)"
… Emily opened this issue a year ago, perhaps Khushal or Victor can speak to this?
Mark: I agree that identity and permission are linked, makes sense to work on them together in an explainer form
Mark: permissions will build on top of agent identity work
Ben: agree, part of agent identity is proof it is acting on behalf of specific user
Anssi: I see Victor made a connection to IETF's work on Web Bot Auth (webbotauth), is this relevant for our discussion?
https://
Yoav: I talked to one of the IETF WG chairs, it seems aimed at server-side communication
… on the client-side, no way to keep private keys secure, seems less relevant here
Anssi: and in context of charter review, Mark made a comment about the agent delegation scope:
<gb> Issue 906 [wg/agenticweb] Agent Delegation deliverable scope could be broadened (by markafoltz)
Anssi: Mark's comment: "The scope of agent delegation should extend across agent interaction modes, including installable agents, and agents running as ordinary web content or cross-origin frames. For agents to be truly useful, they also need other capabilities, such as the ability to observe page content, discover and invoke tools, and simulate UI actions."
Anssi: is this still open?
Mark: I haven't looked at the latest, if this has been addressed, I wanted to soften the language to see what capabilities make sense
… I can follow up on this detail
Mark: I'm referring to the scenarios presented in Dom's slides
<bvandersloot> https://
Benjamin: IETF proposal relevant here is Agent Use of Delegation and Interaction Traceability (audit)
<bvandersloot> Moderation Of unLinkable Endorsement
Yoav: another forming group BOF'd enables bots or agents to pass token acquired by user to prove that they're attached to an actual user rather than bots operating independently
<Ehsan> I think MoLE was for anti-fraud purpose though
<bvandersloot> Porque no los dos?
<yoav> https://
<yoav> some of the identity use cases are very much related to anti-fraud
Johann: my team might be interested in working on Agent Delegation
<Idan> I think the context for MoLE is that many of the new assistants (Muse etc.) are using a cloud VM for all their users, acting as a proxy. traditionally they got blocked and they now need to prove they are acting on behalf of many users
<Mark_Foltz> +1
<domfarolino> +1
<Johann> +1
RESOLUTION: Johann and Mark to draft an explainer for Agent Delegation. (issues #44 and #54)
Agent Interaction Measurement
Anssi: issue #186
<gb> Issue 186 WebMCP "real-user measurement" (by yoavweiss) [beyond-scope: agents-on-the-web]
Anssi: APIs for pages to measure in-field agent interaction performance, reliability
… Yoav shared a strawman proposal building upon the PerformanceObserver spec with new entries for agent interaction measurements
webmachinelearning/
<gb> Issue 186 WebMCP "real-user measurement" (by yoavweiss) [beyond-scope: agents-on-the-web]
Anssi: Yoav, could you talk to your strawman proposal?
Anssi: I think this strawman proposal is a good starting point and would feed into to the explainer
… Yoav, could you take this forward and draft the explainer for Agent Interaction Measurement?
Yoav: PerformanceObserver that monitors tool calls so that would avoid the need to wrap tool calls to count and measure them, also gives us some form of interaction ID to enable us get more visibility into agentic turns
… it might also be interesting to connect this to user interaction
… connect tool calls with UI updates and think how that is to be represented
… tool call is WebMCP specific, agent interaction entries could give us an overview whether tools are used or not, how long user interaction takes from the user's perspective
… it would be interesting to know the outcome of those interactions, user's giving thumbs up or down interaction to agent, agent could assess whether the mission was accomplished or abandoned
Yoav: many entries are optional
Dom: it sounds like this task is extracting information from agent, probably come up with a giant draft and run it through agent vendors to see what metrics would be helpful to them
Yoav: sounds good to me
Dom: the first stab is more like here are the useful metrics
Liad: good idea, in the MCP world we are standardizing intent and feedback, maybe cross-pollinate on these ideas
Yoav: we can outline the use cases and rough skeleton, bits of information we need from agents and review with agent providers
<Vineelaganti> Sai Vineela(Target, enterprise UI/a11y): +1 to measuring outcomes from the user's perspective. Suggest also covering human-oversight signals: did the user intervene, correct or undo an agent action, was a confirmation shown/accepted, and was the agent-caused UI change perceivable (visible update, a11y announcement).
<yoav> +1
<domfarolino> +1
<sjdallst2> +1
RESOLUTION: Yoav to draft an explainer for Agent Interaction Measurement, outline the use cases and rough API sketch, bits of information we need from agents and review with agent providers and privacy group. (issue #186)
Agentic Interaction Modes
Anssi: this deliverable was initially envisioned as a non-normative guidelines doc for agentic interaction in different modes, browser, extension, automation, headless
Anssi: Dom's review comments notes this has normative aspects:
… "a gateway from some defined "agent" into the web platform to normatively expose this from a truthful agent harness."
https://
Anssi: we need to clarify which parts are normative and which are non-normative, and for the normative aspects, have a reference
Yoav: going back to Ben's comments, no rush to go into this particular document, others are higher priority
<domfarolino> +1. It also feels like an extension of "Agent Identity" (not sure if that was already mentinoed)
<domfarolino> _1
<domfarolino> +1
<bvandersloot> +1
RESOLUTION: Merge Agentic Interaction Modes considerations with Agent Identity.
Agentic Web Threat Model
Anssi: I'm pleased to welcome our guest speaker Umar Iqbal, Assistant Professor, WashU
… Umar's research will bring new perspectives into the Agentic Web Threat Model deliverable
… Umar will present novel research on a framework for more secure agentic web interactions, please take it away, Umar
[ Umar presents slides: Securing Agentic Browsing Interactions with Web Semantics ]
Anssi: let's find additional time to get Umar back to present on this research again