Add regular meeting notes from 09.03.2026

Meetings/RegularMeetings/2026-03-09.md

@@ -0,0 +1,80 @@
+# WebAgents CG: Regular Meeting (March 9, 2026)
+
+## Agenda
+
+   * Introduction
+       * Introduction of new participants (if any)
+       * Review minutes from the previous meeting
+       * General CG updates
+   * Interop Report
+       * Arthur Casals: Agent-Environment interaction
+   * AOB
+
+## Participants
+
+   * Andrei Ciortea
+   * Antoine Zimmermann
+   * Jérémy Lemée
+   * Arthur Casals
+   * Olivier Boissier
+   * Andrei Olaru
+   * Ege Korkan
+   * Jean-Paul Calbimonte
+   * Gustavo Nardin
+   * Stephen Cranefield
+
+## Regrets
+
+   * Rem Collier
+
+**Scribe**: Jérémy Lemée
+
+**Notes from previous meetings**:
+
+[https://github.com/w3c-cg/webagents/pull/116](https://github.com/w3c-cg/webagents/pull/116)
+
+## Meeting Notes
+
+New participants: Oliver Boissier
+
+Andrei Ciortea (AC): Proposal of a round table on policies, norms, and accountability
+
+AC: Review of the past meeting:  1) Agent profiles and their importance for Web-based MAS. 2) Paola di Maio presented a video that proposed an ontology for Agentic AI. 
+
+AC: This afternoon [March 9th 2026 4-5pm CET], there will be a meeting of the Web \& AI Interest group. Presentation of our community group. Anyone should be able to join but you need a W3C account [https://www.w3.org/events/meetings/0252457e-6e75-4770-9462-3496ed8e407b/20260309T160000/](https://www.w3.org/events/meetings/0252457e-6e75-4770-9462-3496ed8e407b/20260309T160000/)
+
+AC: Preparing the Interoperability report is a huge task. Proposal: create many smaller reports so that we can publish some parts sooner instead of waiting for the completion of the full report.
+
+### Arthur Casals' presentation: Agent-Environment interaction
+
+The presentation of agent-environment interaction in Web environments starts from the discussion about the Web architecture (section 3.3 in the report). Through hypermedia, agents discover affordances. In description-driven interaction, agents discover machine-readable descriptions to invoke affordances. In protocol-driven interaction, agents rely on standardized interaction protocols for such interactions.
+
+AC: How do hypermedia-driven interaction and description-driven interaction relate? Hypermedia-driven interaction also relies on descriptions.
+
+Arthur Casals: Descriptions-driven interaction relies on "out-of-band or pre-fetched descriptions" (quote from the written section)
+
+Arthur Casals: The section covers different aspects of affordance description, invocation, observability, and context from classical MAS research. After the meeting, the fork with the section of the report will be published online. The first issue concerns (9.1) the identification of environment entities beyond agents and tools (covered in Section 4 of the report). The second issue (9.2) is that LLM-centric protocols collapse the distinction between perception and action. Perception is only the result of a tool call. 
+
+AC: One important question is synchronous vs asynchronous tool use. Issue 9.1 is also a fair point and is required to address to have a coherent discussion across the report.
+
+Arthur Casals: Issue 9.3 concerns the discussion between stateful and stateless tools. Issue 9.4 concerns the action lifecycle in tool-use protocols (e.g., MCP)
+
+Andrei Ciortea: in the Agents \& Artifacts meta-model, artifacts have their own lifecycle and are stateful, while tools for Agentic AI are typically stateless.
+
+Arthur Casals: Issue 9.5 concerns a diagram. Issue 9.6 concerns the normative constraints at the affordance level. It links to section 10. Issue 9.9 concerns information loss in conversion between different standards (e.g., OpenAPI to WoT TD).
+
+Ege Korkan: question about the link between the section and the section on tool profiles. 
+
+Arthur Casals: We need to acknowledge that the problem described in issue 9.9 happens. Issue 9.7 concerns hypermedia discovery in LLM-centric protocols. Can LLM-centric protocols integrate HATEOAS-compatible discovery without being fully redesigned? Issue 9.22: different descriptions model affordances at different levels
+
+Ege Korkan: Point on definition of observability. 
+
+Arthur Casals: Issue 9.19 on tool composition. Issue 9.21 on the scope of environment (e.g., are physical environments considered?)
+
+Jérémy Lemée: questions of the influence of LLM agent architectures (based on ReAct loop) on the proposed issues. Especially, the ReAct loop has synchronous tool use and it is the basis to design LLM agents.
+
+Arthur Casals: A lot of the work on agents/MAS has been ignored in the development of LLM agents/LLM agent systems. 
+
+Andrei Ciortea: The report does not address the agent architectures themselves. However, we can mention what limitations of LLM-based cognitive architectures entail. There is also a lot of development on that domain (e.g., CoALA: [https://arxiv.org/pdf/2309.02427)](https://arxiv.org/pdf/2309.02427)) and the issue may get resolved by the time the report is finalized. 
+
+Olivier Boissier: raises the discussion on reflective use of the environment by agents