Skills Turn Coding Agents into Repeatable Workflow Machines
The shift from 'can AI code?' to 'how to optimize AI coding?' hinges on skills—lightweight, text-based plugins following an open spec, supported by most agents. They inject domain expertise, custom capabilities, and workflows into agents, loaded on-demand via descriptive metadata.
In a demo on the 'Sen' e-commerce site (product pages, reviews), Yohan Lasorsa prompts: 'Implement the first open issue.' The agent uses GitHub CLI skill to fetch the 'add contact page' issue, then builds it. Key skills include:
- GitHub CLI: Describes commands/examples for repo access.
- Playwright: Records videos of changes.
- Custom 'public-tunnel': Creates local tunnel, sends URL via Telegram skill for mobile testing.
An agents.md file orchestrates: After changes, run dev server, tunnel, video record, Telegram notify, hold GitHub issue close until confirmation. Result: Contact page built, video/preview sent to phone in ~minutes, testable remotely. Tradeoff: Agents occasionally fail (e.g., token issues), requiring retries; skills mitigate by standardizing tools over ad-hoc prompts.
"The truth is that today it's mainly a matter of skills but don't get me wrong it's the one that you install and use with your favorite code agent." — Yohan Lasorsa, emphasizing skills as the skill gap for effective agent use.
MCP Servers and DevTools AI Automate Browser Debugging
Manual DevTools workflows (console, network, performance) are now agent-controllable via MCP (Model Control Protocol) servers—tool-hosting endpoints callable by agents from any IDE/CLI. Olivier Leplus demos Chrome DevTools MCP (GitHub repo install via mcp.json):
Tools exposed: click, fill forms, console logs, network requests, Lighthouse audits, navigate, screenshots, resize, throttling.
Prompt: 'Run app, test main page in Chrome.' Agent starts dev server (npm start), launches Chrome (visible automation), screenshots, lists elements. Advanced: Throttle to 2G/3G/fast 3G, trace performance (LCP, CLS, TBT), analyze images/CSS/JS. Outputs report: 'Headphone image too big (preload, high priority), CSS render-blocking, JS preload JSON.' No manual intervention; runs in background.
Built-in Chrome DevTools AI (enable in settings > AI Innovation tab) adds on-device insights:
- Console errors: Click icon for CORS explanation/fix.
- Network 400s: Chat analyzes request context ('endpoint missing').
- Performance traces: LCP breakdown hints ('render-blocking CSS').
- Elements: Live CSS tweaks (e.g., gradient H1 matching CSS vars), 'Apply to Workspace' persists to source files (map workspace folder).
Tradeoffs: AI verbosity varies; DevTools changes ephemeral without workspace mapping. Reduces copy-paste friction vs. external chats.
"What would be amazing if an MCP existed for that like an agent can call it that's what exactly what the Chrome MCP does." — Olivier Leplus, on bridging manual DevTools to agent automation.
Browser-Native Web AI APIs Enable On-Device Features
Cloud AI APIs (tokens, latency, costs) yield to Web AI APIs (W3C draft)—on-device models (~4GB download, cached, auto-evicted on low storage). Demos on Sen reviews:
if ('Summarizer' in window) {
const summarizer = await Summarizer.create({
type: 'key-points', // tl;dr, teaser, headline options; size: words/sentences
inputLang: 'en',
outputLang: ['en'],
context: 'Reviews for headphones. JSON stringify summary of opinions.'
});
monitorDownload((progress) => console.log(progress));
const response = await summarizer.summarize(reviews);
return response;
}
Click 'Summarize': Downloads model once (0-100% instant on cache), outputs: 'Customers praise headphones for sound...'. Similar for Translator, Prompt APIs. Runs locally, no net/costs. Tradeoff: Browser support nascent (Chrome flags); large initial download.
Future: Browsers ship built-in MCP servers; standards like LLMs.txt/WebMCP feed agents docs.
"AI builds the web. The web feeds AI. And now, AI lives inside the browser itself." — Talk intro, framing symbiotic web-AI evolution.
Upgrading Apps for Human + Agent Cohabitation
Agents now 'use' web apps alongside humans, requiring adaptations (teased, demo implied via Sen site). Full cycle: Code/debug/tune with AI, embed native APIs, expose via MCP/standards for agent consumption. No Python monopoly—web democratizes AI.
Key Takeaways
- Install skills (e.g., GitHub CLI, Playwright) in
agent/skills/with descriptiveskill.md; orchestrate inagents.mdfor workflows like auto-tunnel/video/notify. - Add Chrome DevTools MCP to
mcp.jsonfor agent DevTools control: throttling, traces, audits yield actionable reports (e.g., image preloads). - Enable DevTools AI for instant error chats, live CSS/JS tweaks, 'Apply to Workspace' for source persistence.
- Use Web AI APIs (
Summarizer.create) for local summaries/translations; monitor downloads, specifytype/contextfor precision. - Test agent failures early (tokens, sticks); skills standardize over raw prompts.
- Map workspaces in DevTools to avoid ephemeral changes.
- Cache models persist across sites; evict on storage pressure.
- Build for agents: LLMs.txt, WebMCP for doc feeding.
"You've all been through that... you do a lot of changes on the Chrome dev tools and then you don't remember which line you have to copy paste." — Olivier Leplus & Yohan Lasorsa, on CSS tweak pain solved by DevTools AI persistence.