AI Didn’t Kill the Web, It Moved in! — Olivier Leplus (AWS) & Yohan Lasorsa (Microsoft)
Summary
Yohan Lasorsa (Microsoft, Angular GDE) and Olivier (AWS, Web GDE) present a comprehensive tour of how AI has become embedded at every stage of the web development lifecycle in 2026 — from coding agents to in-browser AI APIs to making web apps consumable by autonomous agents.
The session opens with a live coding agent demo built around a sample e-commerce site called "Seine." Yohan demonstrates how Claude Code skills — lightweight, text-format plugins conforming to an open specification supported by most coding agents — dramatically improve agent output quality. Skills stored in the `.agent/skills/` folder are selectively loaded into context based on their descriptions; custom skills shown include GitHub CLI integration for reading open issues, a Playwright CLI skill for recording demo videos, a public tunnel skill for local-to-mobile testing, and a Telegram notification skill. The agent autonomously reads a GitHub issue, implements a contact page, records a Playwright video, opens a public tunnel, and sends a Telegram link — all without manual intervention. The workflow is governed by an `agents.md` file that instructs the agent to record a short video, create a tunnel, and send the URL to a smartphone before closing any GitHub issue.
Olivier then demonstrates the Chrome DevTools MCP server, an open-source MCP implementation that exposes all Chrome DevTools capabilities — screenshots, network requests, Lighthouse audits, performance traces, form interactions, console messages — as callable tools inside any agent IDE. A live demo shows the agent navigating a page, running performance tests across three network conditions (no throttling, fast 3G, slow 3G), measuring LCP, CLS, and render-blocking metrics, then generating optimization recommendations (e.g., large headphone image, CSS blocking, JSON preloading) — all without the developer touching the browser.
Yohan then showcases Chrome's experimental Web AI APIs (Summarizer, Proofreader, Prompt API with multimodal support), currently under W3C draft and available behind flags in Chrome. Key technical details: the Summarizer API downloads a 4 GB on-device model (shared across all websites, cached by Chrome), supports types like key-points, TLDR, teaser, and headline with configurable output length and language. The Proofreader API returns corrected text plus a structured diff with start/end indices for each correction. The Prompt API supports multimodal input (text, image, audio), structured output via JSON Schema constraints, and produces results entirely on-device with no external API calls or token costs. A live demo generates a structured JSON review (title + description) from a photo of a damaged headphone. Opera and Edge have begun implementing these APIs; Chrome requires experimental flags to enable them.
The session then addresses making web apps AI-consumable. LLMs.txt is highlighted as an already-widespread standard — a markdown file at the site root listing links to documentation pages so agents avoid crawling every page. The Angular.dev llms-full.txt variant is shown: a single file containing the complete Angular documentation including code examples, ensuring coding agents use the latest framework version rather than stale training data.
Finally, Olivier demos WebMCP — a highly experimental proposal for exposing MCP tools directly from a webpage. Tools are registered on the `navigator` object using a JSON definition (name, description, input schema, execute function). A browser extension visualizes registered tools and lets the agent call them directly — e.g., "add three laptops to cart" without any DOM scraping. A second pattern converts existing HTML forms into MCP tools by adding HTML attributes (`tool-name`, `tool-description`, `tool-param-description`, `tool-auto-submit`), automatically generating a tool schema from form inputs and labels. This enables agent-ready forms with zero JavaScript changes. The WebMCP API changed 10 days before the talk and the spec is still evolving rapidly.
Key Takeaways
- •Use Claude Code skills (`.agent/skills/`) to give coding agents domain-specific knowledge and repeatable workflows — skills are selectively pulled into context based on their description, so writing precise descriptions is critical.
- •The Chrome DevTools MCP server lets agents run Lighthouse audits, performance traces, and network analysis programmatically — no manual DevTools interaction needed.
- •Chrome's Web AI APIs (Summarizer, Proofreader, Prompt) run a shared 4 GB on-device model with no API costs or token usage — ideal for privacy-sensitive or offline use cases.
- •The Prompt API supports multimodal input (image, audio, text) and JSON Schema output constraints, enabling structured AI responses entirely in the browser.
- •LLMs.txt is already widespread — add one to your site so agents can navigate your documentation without crawling every page; use llms-full.txt to bundle all docs for coding agents needing up-to-date framework references.
- •Register WebMCP tools on the `navigator` object to give agentic browsers direct function-call access to your app's actions rather than forcing agents to simulate human clicks.
- •Existing HTML forms can be upgraded to WebMCP tools with zero JavaScript by adding `tool-name`, `tool-description`, and `tool-param-description` HTML attributes — the browser auto-generates the tool schema from form inputs and labels.
- •Chrome DevTools AI assistance (enable under AI Innovation in DevTools settings) can explain CORS errors, debug failing network requests, and apply CSS changes directly back to source files via workspace sync.
- •Build your `agents.md` to define agent workflow rules (e.g., record a video, send a Telegram notification, don't close issues without confirmation) — most coding agents respect this file.
- Web developers must now optimize for two audiences — humans and AI agents — just as responsive design required optimizing for desktop and mobile simultaneously.
Transcript
Hey folks, welcome to this session where we'll discuss a bit about the web and what the recent AI innovation change for us as web developers. So, let's take just some time for a quick presentation. So, my name is Yohan. I work as a developer advocate at Microsoft. I'm also a GDE for Angular and I'm today with Olivier. Yeah, I'm Olivier, my developer advocate at AWS and I'm also a GDE, but this time on web. So, no surprise here. We'll talk about AI and more specifically, in the last 6 months, the...
Related Sessions
