These Android News Are INSANE - Google IO 2026 Special

Added: 2026-05-25

[00:00:00]Hey guys and welcome back to a new video. There are so many news at this time in the year here to the overall Android and Cotlin ecosystem. There was just Google IO here which this video is all about. There was Cotlin comp with cool changes in the Cotlin ecosystem.

[00:00:14]There is the new Android 17. So many news which is why I decided to actually make two news videos this time. This one here fully dedicated towards all the news from Google IO which tend to be a bit more Android oriented because obviously Android is from Google. and then I will make a separate news episode with all the news from Cotlin Conf that I've this time also attended where you will also get a very nice vlog in the next days or next week probably. So here we really go through all the important things from Google IO. I will focus on those that I consider the most important. So we won't dive into every single detail where maybe 1% of you actually use that but thoroughly dive into parts where a large portion of developers may benefit from that. So here there is an article that summarizes uh the news from Google IO from the official Android developers blog but it's really just a very brief summary of things um and we will dive a little bit deeper here into certain aspects that I consider quite cool or important. So, first of all, of course, Android CLI.

[00:01:10]That's already what I made a video about the new Android command line tool, which is really a great interface nowadays for having an agent, an AI communicate or control your Android Studio IDE, launch emulators, configure emulators, get layout metadata, and so on. I made a separate video about that, so I won't dive into this in this case here, but it was simply part of the announcements in Google IO. But then what is probably one of the most important news to you as Android developers is that Android is now composed first. So what Cotlin did to Java back then for Android, Compose is now doing to the view system. So they are now in maintenance mode which means there are no more features going to be released for the view system. So new features for UI development on Android will only come for compose at this point. Maintenance mode simply means okay if there are of course critical bug fixes for previous viewbased components they will ship these but simply no new features and here really the entire view system is affected. So fragments are affected recycler view constraint layout the navigation framework that worked with the XML nav graphs. So in the end everything that had something to do with the view system is now in maintenance mode. So in case you are living behind a rock and you're still sticking to the view system, then now is really the absolutely latest time for you to switch to compose and learn that because in the end, honestly, it was just a matter of time until this happened. Compose nowadays is on in in a very solid state.

[00:02:39]It wasn't that solid when it was officially marked stable. There were many missing features. You can do pretty much anything with it. There's still some smaller things like scroll bars in lazy columns and so on, but all in all, Compose is just the better framework.

[00:02:52]It's so much easier to build UIs with it. So if you're still sticking to views for some reason, get into compose. But another big chunk of changes are of course how could it else be affecting AIdriven development, affecting agents, affecting new AI tools from Google, and they are actually quite cool this time.

[00:03:10]So some of them I'm I'm a little bit skeptical about, but some others uh are really cool. So where is that here? On the one hand, Google AI Studio. This is one of those news that I'm a bit more skeptical about. So, it's in the end a tool, a browserbased tool uh where you can build native Android apps simply with a prompt in Google AI Studio. This has been around for quite a while already. Uh but now apparently um there have been new features, new changes to this tool. And you can see if you look at it, you can deploy your application on a cloud device. You don't have to boot up your emulator. Everything like the whole development happens here directly uh in the browser. Why am I skeptical about it? Well, if you're honest, if we take a look at that, it looks quite like a a vibe coding tool.

[00:03:54]It's something where you can maybe quickly prototype an app here, quickly experiment with something, but for serious professional development, you of course want all those tools. You want the debugger, you want the profiler, you want to have specific coding conventions that live in your repository. You want multiple developers to be able to work on a codebase. So for for anyone who is either wanting to become a professional developer or is already a professional developer that that doesn't look like something for us but rather for people who say I don't know much about building native Android apps and I just want to quickly get going with something but if you are developing in a professional environment or you want to develop in a professional environment just get into proper agentic workflows that you can execute directly in your IDE be it with cloud code be it with codec any viable agent right now which I've also really um put my best practices for Android architecture in a bundle of eight skills that you can really just install in your cloud code instance or codeex. Codex also works with skills. So whatever you want to use here if you want your agent to stick to my personal Android architectural conventions the you can download the skills here for completely free below. Another piece of news here that I am a little bit skeptical about from the AI space is this one. Convert iOS apps to Android and actually not just iOS apps uh but also you can see React Native web frameworks like Flutter probably um there is a new migration assistant in Android Studio that promises to convert non-Android apps to Android apps. So converting the entire code the UI the framework the SDKs that is how it sounds like. So you can see intelligently map features, convert assets, SVGs, implement Android best practices using Jetack compose and so on. And they actually frame this here to have it transform what used to be weeks of manual porting into an agentic workflow that only takes hours. And yes, maybe it takes hours. So in the sense that the prompt or the actual agent running something takes a couple of hours, this is something that could be realistic. But to have an agent migrate a whole application to another native application in a completely different ecosystem that itself is a great thing and agents are really good at migrating technologies. But I highly doubt that this migration assistancy is really that reliable that you can claim it only takes a couple of hours because proper migration can of course be sped up with agents but it takes longer than a couple of hours because this again gives me this vibe coding feel here that you just say hey you dive into the migration assistant you migrate a whole app to an Android app and then it just suddenly works out of the box. But anyone who has worked in a professional environment knows how much testing is involved when such a migration, such a big refactoring is really happening. And especially if an agent is doing that, you want to be very though with the review. And this is of course technically part of this whole migration. But all in all, it will take longer than a couple of hours. But also to be fair, I have not tried this out.

[00:06:51]If you want me to make a video about that, uh, then say so in the comments.

[00:06:54]But this is really just my feel as someone who has already worked a lot with agents and workflows. Doing this properly is not as trivial as some people want you to believe. But coming to the AI related changes that I'm a bit more excited about that I think are really cool and promising changes are those here about building AI into your apps. So this is now a lot about running certain AI models either on the device so directly on locally in your Android app on the on the phone or being able to connect to cloud-based models from your app. For example, there is this new agent development kit. So the ADK for Android which is in the end an SDK for you to integrate agentic workflows directly into your Android app. So if you have certain AI workflows that maybe process user data step by step sequentially or maybe process some documents and all locally on the device then this ADK really allows you to do that. You can define sub aents. So you can really define multi-step pipelines of agentic workflows of agentic processing after all uh that could then transform certain data from the user that could summarize certain data from the user in a in a multi-step pipeline after all where each sub aent maybe has its own responsibility its own job and this can now actually be implemented directly on the device technically with a complete local LLM model. So also if you have a very sensitive user data that you need to process in some way uh maybe some kind of private documents or so then this way the data will never leave the device and you can still benefit from having an LM maybe summarize such a document maybe transform it maybe extended whatever you want to do then another very cool thing here is and they're actually not diving that deep into this but um I I dig a little bit into these news here in more detail watch some YouTube videos because they are just mentioning one sentence here we also expanded cloud capabilities via Firebase AI logic along allowing developers to leverage Gemini models with robust grounding blah blah blah blah blah. Here this hybrid inference approach this is in the end a hybrid orchestration tool. What does that mean?

[00:08:53]It means that you can have a single entry point for your LLM model in your Android app. So if you want to send a prompt to some kind of AI model from your Android app, then this hybrid inference approach from Firebase will actually intelligently decide whether it should run this prompt locally on the device or whether it should send this to a cloud model which could potentially be more powerful. Also something that they have announced on Google IO that isn't named anywhere here is a very cool API after all for Cotlin data classes to serve as the output for an LLM prompt after all. So what does this mean?

[00:09:28]Normally if you of course prompt an AI and you want the AI to stick to a certain output format for example JSON the AI process some kind of data and you want to map this data into a clear format of maybe JSON key value pairs then in the end that output is still technically a string. it it has a clear format JSON in this case but your Android app would still have to do something with the string would need to send it through some kind of a JSON parser in order to really get this into a cotton data class of data you can work within the app. So imagine you're just building some kind of local LLM here in your Android app and it has the purpose to intelligently search through some documents via a semantic search. So you say give me all the like the top five documents that deal with my health insurance for example along with some metadata about these documents. So where they are placed, what the file path is and so on. And then this output that you want from the LLM after performing the search you of course want to work within your Android app. Ideally, you would like to have a type save list of documents where a document is in the end maybe a data class containing a summary of the document containing the file path maybe containing the the actual line number or in the PDF where it has found something. So some kind of metadata whatever that is. And now what Google has added to their uh AI SDKs in Android is that your LLM model can directly parse this output for those documents into a cotlin data class. So the same thing that happens to JSON structure when we maybe make an API call that responds with JSON then that is also automatically converted into a Cotlin data class with the JSON parsing API.

[00:11:04]The same thing now is the case for output structured output that an LLM responds with that will directly parse this into a cotton data class via annotations. So you simply add annotations to the fields of the cotton data class where you describe in natural language which field represents which part of the LLM's output and then the output of such an LLM prompt can literally be an instance of a cotton data class and this I think is a really really cool API. Also some new cool changes are these two new protocols AGUI and A2 UI. AGUI is in the end a new protocol um that serves for passing agent states and messages. So if you've worked with an agent before then you know that the agent typically executes multiple things in a row in order to just get to the the final result. There may be some kind of tool calls with MCP or so it may execute some commands etc. And when you now have an agent running on a server that goes through such a such a sequence of steps after all and you have a client so an Android app for example then this Android app can now actually get the current agent states via this protocol. So you can directly show in your Android app what the agent is doing, which tool calls it's executing, what the current state is, which calls have maybe failed. So for example, if you have a backend agent running that has the responsibility of actually booking a trip for you, uh then your Android app can now display it, whether that's currently booking the hotel, whether that's currently uh at the the flight booking step for example.

[00:12:28]And this AGUI is in the end just a protocol uh that unifies the communication for an agent to an Android client app. On the other hand, there is this A2 UI protocol which is a way for an AI model to output native UI components instead of raw data. So via this protocol, your Android app would simply tell the agent, tell the AI model, hey, these are the UI components that I support, row, column, button, etc. And for whatever the agent is then doing, the agent presents the best UI that it considers the best UI to the app at that exact moment. So the agent can literally build an Android UI hierarchy for you for something completely customized. Respond with that UI hierarchy to the Android device and the Android device can then display that. So it can display UI that wasn't actually implemented on the device directly that that isn't a fixed composable in your app for example but is actually rendering UI that comes from an agent.

[00:13:22]So taking our uh trip booking agent for as an example again if the agent is for example at the step of booking a flight for you and it notices okay actually the user needs to choose a seat in that flight then that agent could literally just build a dynamic UI at that stage where the user can then select the seat they want to sit at. it responds with that UI and the endote app will simply render it and in and in the end communicate to the agent that way whatever the user has selected for their seat preference and for this they are also really working on a jetpack compose renderer so that uh really composables can be rendered that way from the output from the UI hierarchy that an agent thinks is best and I think these are really interesting changes that are quite buried here in in this uh little blog so you actually need to dig a little bit to find all that out but that's of course what these news episodes are for here another very cool thing regarding AI are these app functions. So app functions is in the end a new API that allows you to take existing functionality from your application. So if you have something like a food delivery app for example that has a checkout flow then obviously you have certain like multiple functions in your app that are executed in order with certain input data that in the end allow the user to perform checkout to make the payment whatever. And now with this app functions API, you can actually bundle such flows into such an app function. And the moment you have bundled that into an app function, it is executable by an agent by an LM running on that device or via a cloud model from Firebase. This means if your app has this functionality of checking out of actually submitting that delivery data to the restaurant, you bundle that in an app function and then maybe have a chatbot in the app and you could tell the chat the user could tell the chatbot please submit my order and the agent would be able to do that with the functionality that you've bundled into the app function. That I think that's super cool because this way you can really make normal functionality of your app accessible to an agent that the user can talk to in natural language. All right, but leaving this topic of AI behind us and taking a look at what else is here. Um, some yeah, adaptive UI, of course, that has changed a lot specifically here for this new Google book. So, that's in the end a new high performance laptop from Google. Um, seems like competing with MacBook. Uh but this also seems to be the reason or one of the reasons why Google is pushing so hard for adaptive layouts in the past years because they probably planned this couple of years ago already and now we actually have an an Android or Google yeah probably an Android powered uh laptop from Google which also comes with a new desktop emulator. So you can now actually have a desktop emulator in Android Studio Canary right now and later in the stable version. Then just some other news that are completely distinct some unified widget development with jetpack glance. So glance is the wi-i widget API uh that you can use along with a jetpack compose. So it seems like it will just be easier for us to develop such widgets be for Android Auto for Android mobile all of that with compose in a more unified way. I'm not sure if there's just one single unified API because at this point um to my knowledge the jetpack lens API uses partly different API. So for example a different type of compose modifier than the um official compose uses for mobile apps. And the way I understand this is that it is going to be a little bit easier for us to for example share UI components between our Android mobile app and such um such a widget to in make it easier for us to build these unified widgets between be OS we're widgets Android Auto Android mobile and also there is a breakthrough they're mentioning um as the integration of remote compose that's in the end the compose server side UI framework uh which is going to be probably a big thing this year where your server in the end can respond with a UI draw instructions in compose to the client device and therefore you can achieve a serverside UI but with native draw instructions. So there is no kind of JSON paring going on or so that the Android device can literally just execute the draw instructions from the server is like a native UI that is however coming from the server and is much more dynamic that way. And other than that there is of course also some yeah some Android for cars changes. I think this only relevant to a fraction of you. Uh so if you're working with Android Auto, you can of course get into that. Here are some media changes. So there are quite some changes to the overall Android media API. Uh post production has been improved. So if you maybe have a video editing app or so um there there are more helpful APIs and functionality now uh in Android 17. But that's also something that I've already covered in the in the past episodes of Android news where uh Android 17 was announced. At least some betas of that have been announced. But this is in the end what I consider the most important changes here from Google IO. If any of that is particularly interesting to you and you say, "Philip, please make a video about topic X from Google IO, about app functions when they're actually publicly available, about building an Android app with an ondevice LLM that maybe processes some documents.

[00:18:22]Let me know that in the comments. I really need to know what you of course want to see here on the channel. And as usual, you can of course find all sorts of super helpful Android and Carlo multiplatform courses in this video's description. Even in the era of AI agents, it is more important than ever to actually understand what they are doing. And that is exactly what you learn in those courses. How you write proper test cases, how you implement CI/CD pipelines, how you really architect an Android app from scratch, how you make system design level decisions, all that is covered in my courses. Check them out in the description. And other than that, thanks for watching. Have an amazing rest of your week. See you back in the next one.

[00:18:58]Bye-bye.