Recap of the web & AI international standards seminar: generative UI and WebSkill

On September 6, 2026, the Web & AI International Standards Seminar on Generative UI and WebSkill, hosted by the W3C Representative Office in Shenzhen and held as part of the 2026 China Computer Federation (CCF) Standards Technical Conference, took place successfully in Shenzhen. The seminar brought together experts from across the industry to discuss emerging technologies and standardization challenges related to Generative UI, WebSkill, and Web AI. Topics included interoperability, performance optimization, accessibility, privacy and security, validation and testing, and how agents integration with existing web platform technologies.

The agentic web era: Technological evolution and governance challenges across web and AI

Chunming Hu (W3C/Beihang) delivered an opening remark [slides, video], observing that AI is evolving from an “answering system” into an “agent capable of taking action,” while the web is shifting from an “information space” into an “action space.” He reviewed how existing web technologies —such as the Semantic Web and WAI-ARIA— can provide semantic foundations for agents, and introduced emerging approaches to web–agent interoperability, including WebMCP and WebSkill. He also highlighted key issues such as agent identity, authorization, and embedding governance into the Web architecture.

Rethinking AI phones and conversational interaction

Vic Yao (Tencent) [slides, video] pointed out that AI phones are not necessarily a disruptive product category and that natural-language interaction has significant limitations. Many smartphone tasks are frequent but relatively simple; their effectiveness can be difficult to quantify; the execution cost of agent-based tasks is high; and natural-language interaction can be ambiguous and less efficient than a graphical user interface (GUI). Generative UI can address some of the limitations of text-based AI output, but it cannot replace our thoughtfully designed graphical interfaces. He emphasized that GUIs and web technologies will remain important in the AI era, and identified WebMCP as a key technology for connecting AI agents with the web.

WebSkill: Building personal AI SaaS services on the web

Chunhui Mo (Huawei) [slides, video] presented WebSkill, a declarative contract designed to enable a complete personal AI experience to run within the browser. WebSkill is defined by three core properties: self-contained execution, strong privacy isolation, and dynamic evolution with personalized assets. Skill files are stored in Origin Private File System (OPFS), while scripts execute inside a secure Web Worker sandbox, ensuring sensitive data remains entirely client-side. He demonstrated a range of use cases, from creating Skills using natural language to generating project briefings with a single action and analyzing quality issues through natural-language instructions.

Applying generative UI in search

Jiexuan Gao (Baidu) [slides, video] shared the production practices of generative UI in search across two phases: Phase 1 leverages a custom DSL (Markdown extensions + UIJSON protocols) for stable outputs, to work with the models’ early-stage capabilities; Phase 2 constructs an agent pipeline for direct HTML rendering, utilizing a harness system to enable a constrained, verifiable, and evaluable production process. The key lies in establishing protocols and standards and reusing expert assets, rather than relying solely on the model.

Towards a quality assessment framework for generative UI

Jiadong Yan (Huawei) presented a quality evaluation framework for Generative UI, pointing out challenges such as multi-modal coupling, heavy reliance on rendering, open-ended goals, and high subjectivity. He proposed breaking down the user journey into a layered design consisting of a visual layer and a source code layer. Furthermore, he distilled key experience metrics, including display correctness, requirement satisfaction, functional consistency, end-to-end success rate, timeliness, and esthetics, and shared their internal experience with applying this framework.

AI, web and world

Yipeng Zhao (Tencent) [slides, video] shared his insights on “Web and the 3D Digital World in the AI Era.” He pointed out that while 3D AIGC and 3D reconstruction technologies are developing rapidly with explosive production capacities, they lack connectivity, making it difficult for them to truly integrate into daily life. To address this, he proposed drawing on proven web concepts, including sites, hyperlinks, HTML/CSS, and browsers, to architect an open, interoperable 3D digital ecosystem in a lightweight, rapid, and ubiquitous manner. Ultimately, this approach aims to facilitate AI-driven applications such as web gaming, 4DGS dynamic performances, and AI-powered NPCs.

Open discussion and next steps

Li Li (Huawei) [video] kicked off the open discussion session and raised a series of questions that sparked lively discussion among the participants: what work and technology directions can be explored; what standards can be developed; and how the broader business model (such as the current attention economy and advertising model) might change in the future.

Participants expressed strong interest in continuing discussions within the W3C community on topics including the incubation of the WebSkill proposal, a quality monitoring and evaluation framework for generative UI, and the role of the web in AIGC worlds. 

As a next step, W3C encourages speakers and participants to continue these discussions through breakout sessions at TPAC, Community Group discussions, and other community activities. As the agentic web era arrives, the W3C community’s core focus will be on a key challenge: how the web platform can better serve the joint needs of humans and agents without compromising on openness, security, or accessibility.