Like every year, I binge-watched I/O. And like every year I came out the other side with more links than time. A fire-hose of announcements, most of which you’ll never touch, a handful of which will change how you work next week. Here I filter for you. Links to the sessions are inline so you can go deep where it applies to you.
The keynote, briefly
I’ll keep this short because you’ve probably already read it somewhere (or watched it!).
Gemini 3.5 Flash landed. Cheaper, very fast, and comparable to Opus, which is a bold thing to put on a slide. Four times faster, and apparently twelve times faster in Antigravity IDE. I wrote “how?” in my notes at that point and I still don’t have an answer of how an IDE can make it 3x faster. 3.5 Pro follows next month.
Gemini Spark is the personal-AI play. Not in Chrome yet, that’s this year and not on Android until next year. Rolling out over the next few weeks, with MCP support for non-Google tools.
Generative UI via Antigravity is now in Google Search. You search, it writes code to build the visuals, so you get a unique response rather than ten blue links. Whatever you think of that, it’s a genuinely different shape of product.
Then two protocols that got less stage time than they deserved: the Universal Commerce Protocol for agents selling products, and the Agent Payments Protocol for agents making payments on your behalf. It’s plumbing that decides whether agentic commerce is a real thing or a demo reel. The interesting shift this year isn’t agents that chat; it’s agents that transact.
Gemini Omni is the umbrella: create anything from any input, “starting with video”. Underneath it sits Nano Banana for images, Veo for video and Lyria for music. And there was the obligatory JAX and Pathways slide about the largest training cluster in the world — impressive, not actionable.
App Functions: the one to act on now
If you watch one Android session, make it What’s New In Android (duh).
The headline for app developers is App Functions, now in preview. It lets you declare what is effectively an MCP surface for your app, so Gemini can call into it. There’s an App Functions Jetpack library and a test agent to go with it. Early access sign-up is at goo.gl/eap-af.
I’d say this is the single most important thing to act on from this year’s I/O. Being callable by an assistant is the new deep-link — and the apps that get their functions declared early are the ones that end up in the answer rather than underneath it. Get on the list.
The same session had a grab-bag of other things worth knowing about:
- A2UI — a Jetpack Compose renderer.
- New photo, contact and eye-dropper pickers.
- Pointer remotes on Android TV, and the CAL (Car Apps Library) parked video experience — “hd video whilst parked”.
- GoogleBook, Google’s new laptop, pitched around “desktop class apps”.
And then the strategic one: Compose First, with views now officially in “maintenance mode”. If you’ve been fighting that argument with your team about XML, here’s another reference to drop.
Android Studio: the tooling got a ++
What’s New in Android Developer Tools was the session I got the most out of, by some distance. There’s a lot in it, so let me group it.
Agent ergonomics. You can run Gemma locally and offline now, and the new IDE ships with models built in. But the clever bit, the cleverest idea in the whole session, I think, is that Android Studio catches sed, awk and grep calls from the agent and translates them into IDE commands. That means static analysis actually runs over the change, and the AI gets that feedback and corrects itself. The demo was a Uri extension function. Closing the feedback loop like that is what separates a useful agent from a fast but not so powerful one.
Alongside that, all “safe” commands are automatically allowed; no allowlist babysitting, no clicking approve on git status for the fiftieth time (the direction of all LLM CLI’s tbh). And the android CLI’s emulator start command now waits for the emulator to actually be up rather than returning immediately, so you can pipe it into whatever comes next. Small thing. Enormous quality-of-life win.
Code review, reimagined. Reviews are no longer file-by-file in filesystem order. There’s a review tool now; Git View, then the Review button, that takes the diff chunks, regroups them logically and sorts them into an order that makes sense. Reviewing a change should read like an argument, not a directory listing, and this is the first tooling I’ve seen that treats it that way.
Migration tools. Two of them: one that converts an iOS or React app into Android, and one that migrates XML to Compose. That second one pairs directly with the “views in maintenance mode” line from the other session. Converting XML to Compose is sometimes a genuinely hard problem and I’d temper expectations accordingly — but having it in the box changes the calculus on old screens.
Performance and correctness. There’s a new system trace tool that captures screenshots to help you orient yourself in the trace, which is a lovely touch if you’ve ever squinted at an unlabelled timeline. There’s an agent skill for it too, so you can literally ask “why is my startup so slow, it’s over one second?”. LeakCanary is in Android Studio now, grab a heap dump and point an agent at it. Versus LeakCanary’s own Shark tool, which offers the same thing on the CLI to any agent you like. And there’s an R8 configuration analyzer that shows not just your own keep rules but every library’s, what classes they keep, what percentage, and lets you drill in.
LightBuild got a demo here too. Fully declarative, all data. The pitch is that Gradle sync is essentially a dry-run mode you have to execute to discover everything about your build, and a fully declarative build doesn’t need it. Worth reporting the framing straight: it’s “not meant for pro developers, meant for simple… agentic vibed apps”. GDE’s have closed preview access and they’re hoping for open preview in “weeks”. Even if you’re not the target audience, it’s worth watching what a genuinely declarative build implies.
On-device AI is ready for real work
Deploy Android on-device AI with ML Kit GenAI and LiteRT-LM is the session to watch if you’ve been curious but haven’t started.
The why is straightforward: processing prompts and data directly on device gets you user privacy (banking, medical records, the stuff that should never leave the handset), offline availability, and no cloud cost. The where is equally clear, short content and short context windows. Classifying content, translating short sentences, summarisation, proof reading, rewriting, image description, speech recognition.
There are two routes in. ML Kit GenAI APIs, where Google manages the model for you through the AI Core system service, or LiteRT-LM, where you manage the model yourself. Useful detail if you’re prototyping: prompts written for Gemma 4 work for Gemini Nano, because they share an architecture.
Two practical things I noted down. Prefix caching: if your prompt always starts with the same block of text, caching it makes inference roughly twice as fast. And structured output is coming, an API that guarantees the model returns a data object, which you can then type straight to a Kotlin data class. That second one quietly removes the most annoying part of shipping LLM features.
For resources there’s the Gemma Cookbook for training your own models, the AI Edge Gallery app on the Play Store for inspiration and examples, and the Google AI Edge Portal, a hundred physical devices to benchmark your model across before you ship. This is the unglamorous option that saves you a production incident, so don’t skip past it.
The niche gem
Build a modern note-taking app on Android — an entire I/O session, for note-taking apps. And yet I’d tell you to watch it.
The latest SDK for brush strokes is well documented and abstracts away all the low-level detail and the memory problems that come with it. But the thing that made me sit up is Brush Designer: a visual tool where you draw your brush stroke, tune pressure and tilt curves, and it hands you the code to reproduce it. You get to it from the settings screen of the Cahir sample app on GitHub. Stroke speed can drive opacity, so a flick of the stylus makes the pen line fade out at the end. It exports a .brush file, which is a gzip-compressed protobuf. Very cool.
There’s also a bit in there on multi-window and the desktop experience, which is more broadly relevant than the title suggests.
Two more worth your time
The fireside chat had the framing that stuck with me longest: technical debt, cognitive debt and intent debt, and Margaret-Anne Storey’s point about how AI can accelerate making all three of them better and worse. That’s the honest version of the AI conversation and it’s worth twenty minutes of your attention.
And Build core skills to thrive as an AI-era developer is on my list of ones to point people at.
So what should you actually do?
If you watch nothing else, watch the developer tools session, that’s where the changes to your daily loop are. And whatever else you do this year, get on the App Functions early access list at goo.gl/eap-af, because making your app callable is going to matter sooner than it feels like it should.
If it’s put you in the mood to build something, there’s the Build with Gemini XPRIZE Hackathon over at geminixprize.com.
Thanks for reading. Happy coding!

Leave a Reply