Michael Hablich

@testy_cool @ChromiumDev a few things: 1.) the audit should be applicable for any significant browsing agent out there 2.) SEO is another topic anyways - but related. 3.) audit is experimental and will change, file feedback on our repo please.

@ChromiumDev Oh dear, so now we finally care about /llms.txt? developer.chrome.com/docs/lighthous…

Prepare your site for AI agent interaction with Lighthouse → goo.gle/3ReK3JD If you want AI agents to actually navigate your site properly, the new experimental audit in Lighthouse lets you see: ☀️ Discoverability for AI agents ⚡ WebMCP integration 👀 AI accessibility #GoogleIO

@C1B_nb @ChromiumDev eventually yes.

@ChromiumDev So pagespeed.web.dev will show it now!?

@ChromeDevTools for agents is officially out of preview and now generally available! More infos: developer.chrome.com/docs/devtools/…

That’s a first: GitHub’s Advisory Database (incorrectly) flags Puppeteer as malware. github.com/github/advisor… We are no longer able to publish new releases. Anyone at @github or @Microsoft who can escalate this, please?

We are entering the era of dual interface products: - Rich UI for humans - Token efficient API for agents Your API is costing agents money. I'm speaking at Agentic DevDays in Stockholm on April 28 to share how we are addressing this for @ChromeDevTools .

Want to give your agent quality checks? Chrome's DevTools MCP now includes: ⚡️ Performance checks via Lighthouse 📈 Memory leak detection Skill 🦻 Accessibility debugging Skill 🎨 LCP optimization Skill and an experimental new CLI 👀

@ThoBustos Thanks for the pics!

@MHablich on building agents interfaces

Day 3 starting with Gemma 4. @aiDotEngineer

@qtnx_ @shijith @piniwit @b_roziere github.com/ChromeDevTools…

@MHablich @shijith @piniwit @b_roziere i see, sorry read a bit too fast, for the doc in chrome devtools this is how to add it to vibe add in ~/.vibe/config.toml [[mcp_servers]] name = "chrome-devtools" transport = "stdio" command = "npx" args = ["chrome-devtools-mcp@latest"] if you want to discuss more…

@mistralvibe I would love to get working setup instructions for Mistral Vibe added to #mcp-client-configuration" target="_blank" rel="nofollow noopener">github.com/ChromeDevTools… ... want to open a PR and cc me please?

@shijith @qtnx_ @piniwit @b_roziere Exactly :-)

@qtnx_ @MHablich @piniwit @b_roziere oh think there's been a misunderstanding here. Michael takes care of chrome devtools mcp for google. And there are no instructions on their github page on how devtools mcp can be added to mistral vibe. So michael was just looking to connect with someone at Mistral to correct that

@shijith @piniwit @qtnx_ @b_roziere Sadly radio silence :-/

@MHablich Dont know if anyone has gotten in touch with you yet. If not, hopefully one of these guys from Mistral will. CC-ing @piniwit @qtnx_ @b_roziere

@MistralAI gentle ping 😺

Oh, that's going to be interesting (in a good way)

Your #AI agent isn't picking the wrong tool because it's dumb. It's picking the wrong tool because your description didn't help it pick the right one. Link: hablich.dev/articles/disco…

Wrote about something that's been on my mind: every browser #AI agent tool runs on CDP, which has zero auth. The security story is entirely in the layers around it. Checkout hablich.dev/articles/trust…

@kurtextrem True and there will be a gravitational pull towards more optimized interfaces like WebMCP. I wonder if screenreaders/accessibility tools will also use WebMCP to do actions on a page, instead of trying to navigate the a11y tree.

@MHablich Kinda related: a11y is now becoming more important as agents interact with the accessibility tree. Good thoughts!

1. We spent a decade building web performance metrics around human perception. Turns out they matter just as much when your user is an AI agent. The reasons are completely different. 🧵

7. If you had to set a performance budget for agent users, what would you put in it? #WebPerf #AIagents

6. The mental models still hold: set a budget, measure, optimize the critical path. But the budgets change, the measurements change, and what counts as "critical" changes when the user doesn't have eyes.