Google I/O 2026, Day 2: When the Hype Meets Your Git Repo (Part 2)

Day one was Gemini eating the world; day two was “ok, but how do I ship something before Q4?” The keynote glow was gone, and it was time for CLIs, SDKs, migrations, and all the stuff that breaks on Friday evening.

Table of Content

• From fireworks to developer mode
• Antigravity grows from copilot to platform
• Android 17: your codebase gets an agent too
• Web dev gets its own AI spec: WebMCP
• Modern Web Guidance and HTML-in-Canvas: because your agent also writes CSS
• Chrome DevTools for agents: your bot gets a debugger
• So, what’s left after the smoke clears?

From fireworks to developer mode

Google spent the second day doubling down on one idea: agents are not a demo, they’re supposed to sit in your toolchain and grind. Most of the bigger announcements landed in the developer keynote and follow-up sessions, with a clear focus on Android, web, and Antigravity.

The agenda was packed with “What’s new in Android” and adaptive development sessions for phones, cars, TVs, and XR, all orbiting around Android 17. And starting May 21, over 85 sessions and codelabs are available on demand if you feel like binge-learning instead of touching grass.

Antigravity grows from copilot to platform

On day one, Antigravity 2.0 was the shiny “agent-first desktop app.” Day two clarified what that actually means for your workflow.

• Antigravity CLI is now available to all Gemini CLI users, bringing the same agent harness to your terminal.
• You can spin up specialized sub-agents to handle long workflows, with sandboxed terminals, credential masking, and hardened Git policies built in.
• Managed Agents in the Gemini API give you a fully provisioned agent with a remote sandbox from a single API call, so you don’t have to babysit infra.
• If you hate managed services on principle, there’s an Antigravity SDK to host the agent harness on your own infra and wire in custom tools and policies.

In short: yesterday it looked like a fancy AI IDE. Today it looks suspiciously like the runtime for “agentic backends” you’ll be asked to evaluate in your next sprint.

Android 17: your codebase gets an agent too

The Android part was very clear: Gemini and agents are no longer just user-facing toys, they’re inside the dev tools.

• A stable Android CLI lets agents do the grunt work Android Studio used to nag you about: download SDKs, run apps on devices, and manage project plumbing.
• Google open-sourced Android skills so LLMs know how to handle annoying tasks like migrating to Jetpack Compose or updating to Jetpack Navigation 3.
• Android Bench is a leaderboard for Android development tasks where models, including open-weight ones like Gemma 4, get scored on real Android problems.
• A new migration agent in Android Studio can convert existing apps to native Kotlin Android, whether the source is React Native, a web framework, or even iOS.

That last one is spicy: migrations that used to take weeks are now pitched as “hours”, with the agent doing the bulk of the code rewriting and you doing the swearing during review. The message is obvious: if you’re still procrastinating your Compose rewrite, there is now an AI that will happily mess up your code faster than you ever could.

Web dev gets its own AI spec: WebMCP

Chrome and the web platform also got the “agentic” treatment, but in a way that actually makes sense if you build sites for a living.

The headline piece is WebMCP, a proposed open web standard for exposing structured tools - JS functions, HTML forms, and more - to browser-based AI agents. The goal is to let agents operate on your site with real actions instead of scraping like a desperate Selenium script.

• WebMCP enters origin trial in Chrome 149, with Gemini-powered agents in Chrome coming soon.
• You expose capabilities as tools, and agents can call them with proper structure and error handling.

This is basically “turn my site into an API for agents” with a spec instead of a thousand hacks. Also a convenient excuse when PMs ask for “full AI integration” and you just want a clear standard to point at.

Modern Web Guidance and HTML-in-Canvas: because your agent also writes CSS

To avoid agents reinventing bad patterns, Google is shipping Modern Web Guidance, a set of expert-vetted skills for AI coding tools. Think of it as a giant “don’t do that” pack for performance, accessibility, and security.

• Launching in early preview, it covers over 100 use cases.
• Integrates with Baseline, so it knows what features are actually safe for your target browsers and suggests fallbacks.
• Installable in one click from Antigravity or via CLI with npx modern-web-guidance install.

On the more experimental side, there’s HTML-in-Canvas, a new declarative API that lets you mix real DOM elements inside a WebGL or WebGPU canvas. It’s available as an origin trial, and the promise is “immersive 3D, still searchable and accessible, still friendly to built-in browser features”.

So yes, they’re trying to let you build flashy app-like experiences without fully giving up semantics and accessibility… again. This time with an AI in the loop, which will absolutely propose putting 17 buttons in one canvas.

Chrome DevTools for agents: your bot gets a debugger

One neat bit that didn’t get as many headlines: Chrome DevTools for agents. Instead of you staring at Lighthouse reports at 2 a.m., your agent can now do the boring verification step.

• Agents can run quality audits, emulate real-world user scenarios, and debug layout issues in real time.
• Sessions can be handed over with auto-connect, so you can jump into an active agent session when things inevitably go weird.
• It closes the loop: the same agent that writes the code can now see how it performs at runtime and iterate without manual babysitting.

Basically, your CI/CD is about to get a loud new coworker that files tickets instead of memes.

So, what’s left after the smoke clears?

By the end of day two, the pattern was obvious: day one sold the agentic Gemini era, day two stuck it into your editors, CLIs, and browser tools. Android 17, Chrome, and the web stack all now come with an assumption that there’s at least one agent sitting between your commits and production.

Whether that’s a dream or a nightmare depends on how allergic you are to “AI migration agents” touching your legacy code. But one thing’s clear: if you ignore this wave, it won’t be for lack of APIs, SDKs, or CLIs - Google just spent two days shipping those directly into your workflow.