André Bandarra

Happy World Browser Day! I've long used "long live the web" as my motto/tagline, so today I published a blog post explaining what this phrase means to me: longlivetheweb.org/not-a-manife...

Microsoft Edge's #WebMCP implementation is ready for testing, and the team has new demos! Can't get enough pizza related WebMCP demos! 🍕🍕🍕

WebMCP landed in ChatGPT today Websites expose tools, your agent can use them, or you can just use the app like a human. Probably both! I'm calling it Clicks-n-clankers™

🛠️ I've made a useWebMCPTool hook for React! Chrome (myself and others) will keep this up to date with any spec changes. The hook has hygiene around unmount, result normalization, etc. Cheers! www.npmjs.com/package/use-... github.com/GoogleChrome...

AI agents are changing how users interact with the web. Is your site ready? 🤖🌐 To participate in the Agentic Web you need "agent-readiness." Focus on Discovery, Understanding, and Usage to make your site accessible for AI. Read more at bandarra.me/posts/the-ag... #AgenticWeb #WebMCP #AI

I've built an extension for Pi that lets you communicate with a running instance of pi via discord, telegram, email and web hooks I'd love some early testers to help me with on-boarding and testing. You can reply or DM and I can send you the link.

Really insightful workshop about WebMCP at #GoogleIOConnect - this actually might make a lot of sense to integrate in our appointment booking product. What makes it useful is that agents do not have to use the same flows as people and can answer questions like: “when is the next available time?”

Whether your are building an AI Agent that uses WebMCP tools or exposing tools on your site to AI Agents, there's security guidance to take in consideration! #WebMCP #AgenticWeb

Are you building with Angular and wondering how use WebMCP to expose tools to AI Agents? The Angular team has got you covered and published detailed docs on how to use WebMCP with the framework! Read all about it at: angular.dev/ai/webmcp #WebMCP #AgenticWeb

Testing generative AI is difficult when one input leads to thousands of unpredictable answers. We published a guide on WebMCP evals to help you test LLM touchpoints. Verify that agents understand your schemas and can successfully complete full user journeys → goo.gle/4ftQUZI #GoogleIO

@tomayac.com 's I/O talk has some great ideas for incorporating AI features into a web app, using the Built-in AI APIs. I'll definitely steal some of those ideas and put them in the admin interface for my personal blog Watch the talk here: www.youtube.com/watch?v=ddBx... #WebAI #BuiltinAI

@kulikowski.bsky.social 's I/O talk is a treasure trove to prepare your site for AI agents: 1. How agents 'see' your site, best practices to prepare, and developer tooling. 2. WebMCP deep dive with cool demos, Puppeteer & DevTools support. Watch: www.youtube.com/watch?v=HdCc... #AgenticWeb #WebMCP

🛠️ Ultimately, WebMCP is about making a more secure agent pathway instead of scraping DOM and a11y trees. I made a small site to demo how it works, how agents use it, what the spec looks like, and some important links. It's interactive! You can use the extension to actuate the page:

The Prompt API is now in stable with Chrome 148. Use it with Gemini Nano to build web AI experiences with multimodal inputs and structured output for richer experiences. Plus, you'll now have access to expanded language support → goo.gle/43hL06y

🚀 Google I/O updates for Chrome built-in AI! 🎉 Prompt API ships to stable in Chrome 148! Previews on the horizon: 🛠️ Native Tool Calling preview 🔍 Embeddings API preview. Also a new 🎛️ Prompt Parameters Origin Trial. Join the Early Preview Program for heads up and feedback: goo.gle/chrome-ai-de...

Thrilled this is out, keeping quiet has been SO HARD! 🤐 I’ve had the opportunity to contribute as a domain expert consultant over the past few months, and saw firsthand the level of thoughtfulness and care that went into it. Proud to have played a small part, and excited to see what folks build! ❤️

WebMCP will enter the Origin Trial in Chrome 149, just a few weeks away. But you can try it today by enabling a Chrome flag! Want to now more? Check out the docs at developer.chrome.com/docs/ai/webmcp

I wrote a Markdown editor on the web that has an AI Agent attached to it. The agent an read, write, edit, read other documents in the workspace, delegate to other agents and more. Check it out at bandarra.me/apps/agent-t... #AgenticWeb #AI

Zack's demo (from @brilliantsole.bsky.social ) got my week off to a great start! 🚀 Pair smart glasses with the Prompt API and a few other technologies, and apparently, you get a digital first mate with a voice of pure authority. 🏴‍☠️ ​Check the vibe: youtu.be/tcqrFwadepQ?...

It's time to wire everything together into a continuous testing pipeline → goo.gle/4tSIcZo Today, we'll cover extended unit tests for PR merges, weighted Composite Metrics for nightly integration runs, and Pairwise Evaluation to mathematically prove which model is best when you want to swap LLMs.

AI agents often feel bolted to sites. Moving the agent loop to the client-side enables tight integration between the agent and your site's user interface, while still allowing usage of big cloud LLMs or delegating tasks to server-side agents Read more at bandarra.me/posts/bringi... #AI #AgenticWeb

Great web architecture benefits everyone, from users to AI agents → goo.gle/3Obk1pn Revisit the core tenets of accessible and semantic HTML to ensure your content is discoverable and easy to navigate for all.

Building reliable AI tests require more than asking an LLM for a 1-10 score. By combining objective checks with an LLM judge that outputs a rationale and then a binary score, you get a much more robust pipeline. Learn the steps to set this up in part 4 and 5 of our AI evals series → goo.gle/4uDaaZl

Building on the previous video, can AI write code to solve the maze? It turns out it can! Check the video and read more at bandarra.me/posts/webmcp... #WebMCP

Can we solve AI agent latency by changing where logic executes? 🔌 I'm investigating client-side code execution via #WebMCP. Instead of high-latency tool round-trips, the agent generates a local JS algorithm. Great for speed, but security trade-offs remain. Details: bandarra.me/posts/webmcp...

Learn the science of AI evaluation with our new online series featuring @maudnals.bsky.social → goo.gle/4tOVshb We're sharing a scientific approach to building eval pipelines that ensure your app actually works for users 🧪

Is your AI agent truly "magical," or just a `while` loop with good PR? 🤖 To build production-grade agents, you have to trade magic for mechanics: 🧠 Brain (LLM) 🛠 Tools 💾 Memory 🔄 Engine Read about the plumbing behind these systems: bandarra.me/posts/demyst... 🦀

This is a call for your creativity! 🫵 Everyone seems to be so excited about HTML-in-Canvas, and understandably so. Have you seen all the impressive demos developers have been sharing? We've started collecting some examples: github.com/GoogleChrome... Add your own demos!🎨 🎉

Try the Built-in AI extension to get access to browser APIs like… 🧠 `LanguageModel` 🗜️ `Summarizer` ✍️ `Writer` 📝 `Rewriter` 💡 `LanguageDetector` 🗯️ `Translator`

New AI UI spotted: Readymade categories for user feedback on AI experience (= failure cases) Source: kayak.ai by KAYAK I'm scouting new UX/UI patterns I find in the wild as more AI ships to production, for science 🧪 ➕ Follow for more #UX #UI #AI #appliedAI

I deconstructed an AI Alt-Text demo to move from AI as a detour to AI as a "Main quest" helper. Think "Push-bars" instead of "Pull" handles: don't force new habits; meet users where they already go. Demo: exploractical.com/demos/alt-text/ Underlying thinking: exploractical.com/blog/2026/alt-text/

Been exploring #WebMCP — it only gives agents tools, no prompts, no injected context. I tried: tool description = hint, return value = full protocol. The agent calls the skill and gets the instructions. Lazy-loaded knowledge. Full write up: bandarra.me/posts/webmcp...

Can the AI agent escape the maze? In this video, I show how an agent can go from using direct commands that trigger WebMCP tools to move around the maze, to how the agent can combine the tools provided to solve the maze without having to guide the it step-by-step.