Android Feed

← The 5 stories that mattered

Everything collected ·

All 403 items collected for this issue

New here? How to read this issue
  1. Read the picks. Each tab opens with 4–5 stories ranked by how much you’d regret missing them. That takes about 3 minutes. “Start the read” shows one story at a time, and “Listen” reads them aloud.
  2. Stop at “You’re caught up”, or go deeper. Everything after it is the full list of what was collected today, folded away.
  3. Android comes first. Product & AI covers what builders are launching and discussing.
Stable · Beta · Alpha
Release stages. Stable is safe to ship, beta’s API is frozen but may still have bugs, alpha is experimental.
AndroidX
Google’s Jetpack libraries (Compose, Paging, Room…), released separately from Android itself.
AOSP insider
Code changes merged into Android’s open source repository. They often show an API before it’s announced.
Lanes
Where an Android item comes from: official, AOSP insider, experts, open-source library releases, leaks and ecosystem, community.
Adopt now · Watch · Rumour
What to do about it, and how sure the source is.
Dotted underline
An inferred claim: Claude’s reading, not something the source says.

Collected automatically each morning from 37 sources. Notes are written by Claude. Links show what was trending, not endorsements.

Android5Product & AI4
Start the Android read5 stories, ranked by what you’d regret missing · 2 min read

  1. Grab cuts ANRs by a third with R8 and startup profiles

    If your app ships without R8 optimization or baseline profiles, this is one of the cheapest speed and stability wins left on Android.

    Google's Android Developers Blog published a case study on Grab, the Southeast Asian superapp for rides, food delivery and payments. Grab used R8, the tool that shrinks and optimizes app code, together with profile guided optimization, which prepares the code users run most. The result was about 32 percent fewer ANRs, the app not responding errors, and about 35 percent faster startup.

    Android Developers Blog · OfficialRead the sourceProduct notes

  2. One malformed text message can crash Android's phone service

    A text message that can end an emergency call is a platform bug, so teams building telephony, SMS or system apps should watch for the fix.

    A report that reached the Hacker News front page and Lobsters says a single malformed WAP Push message, a kind of SMS used for things like carrier settings, crashes com.android.phone. That is the system process that runs calling. When it crashes, active calls drop, including emergency calls. The Hacker News thread drew about 180 comments.

    Lobsters · android · CommunityRead the sourceDiscussion

  3. DataStore 1.3.0 is stable, alongside WebKit 1.18 and Palette 1.1

    Stable versions are the ones safe to ship, so this is a good week to update DataStore and WebKit in apps that use them.

    Google released version 1.3.0 of DataStore as stable. DataStore is the Jetpack library that replaces SharedPreferences for saving small settings and typed data. The same release batch made WebKit 1.18.0, the helper library for WebView features, and Palette 1.1.0, which pulls colors from images, stable as well.

    AndroidX releases · OfficialRead the source

  4. Android 11 still runs on about one phone in ten

    Before raising your minimum SDK, check your own users, because one phone in ten on Android 11 means old versions still matter.

    Android Authority reports that Android 17, this year's release, is barely making a dent in the numbers, while Android 11, released in 2020, still powers about 10 percent of phones. More than a quarter of users have upgraded, but millions of devices are still on outdated software.

    Android Authority · EcosystemRead the source

  5. Why coroutine tests flake, and the habits that stop it

    Flaky tests cost a team more trust than failing ones, and this post targets the dispatcher and scheduler habits behind most coroutine flakes.

    Kevin Schildhorn at Touchlab wrote a guide to hardening Kotlin coroutine tests, and Android Weekly picked it up. Coroutines are Kotlin's way of writing code that waits without blocking. The post explains why tests that use them pass on one run and fail on the next, and which habits around test dispatchers and schedulers remove the flakiness.

    Touchlab (Kotlin Multiplatform) · ExpertRead the source

Also worth knowing

Releases, policy and leaks an Android engineer should know about this week.

Go deeper: every Android item, by lane

Official & releases

56DataStore 1.3.0, WebKit 1.18.0 and Palette 1.1.0 are stable and safe to adopt, Android Studio Rabbit 3 Canary 2 is out, and Google's Grab case study shows R8 and profile-guided optimization cutting ANRs by about a third.

AndroidX library releases

Stable · safe to ship · 8

Beta · API frozen, try it before it ships · 13

Alpha · experimental, APIs may change · 23
Show 4 more
  • Android Studio Rabbit 3 Canary 1 now available

    Android Studio Rabbit 3 Canary 1 is now available in the Canary channel. If you already have an Android Studio build on the Canary channel , you can get the update by clicking Help > Check for Updates (or Android Studio > Check for Updates on macOS). Otherwise, you can download…

  • Android Studio Rabbit 2 RC 1 now available

    Android Studio Rabbit 2 RC 1 is now available in the Beta channel. If you already have an Android Studio build on the Beta channel , you can get the update by clicking Help > Check for Updates (or Android Studio > Check for Updates on macOS). Otherwise, you can download it here…

  • 1.13.0-beta01

    Changes since 1.13.0-alpha02 Migration Notes Desktop Libraries using experimental DragAndDropTransferAction must be recompiled because of a binary incompatible change #3413 Features Multiple Platforms Support for rendering Bullet annotations in AnnotatedString #3400 iOS Add…

  • Smaller apps and faster GPU inference with LiteRT in Google Play services

    LiteRT in Google Play services adds about 1.4x faster GPU inference and a CompiledModel API.

AOSP / AndroidX insider

27Gerrit removes the abandoned Privacy Sandbox ads library, adds an annotation toolbar to the Compose PDF viewer, builds an unreleased Information Matrix component, and puts quantized scoring for on-device app search behind a flag.
Show 19 more
  • Annotation processing: detect unreadable annotationsnew

    Add XPropertyElement.hasInaccessibleAnnotations()

    XProcessing, the layer under Room and Hilt's processors, learns when annotations cannot be read, which should mean clearer errors instead of silent misses.

  • Internal settings for one-handed gesturesnew

    Add an internal one-handed gesture configuration

    A configuration object to go with the one-handed gesture modifier, so apps could tune or disable gestures once it goes public.

  • Information Matrix: slot components addednew

    Add the Information Matrix slot components

    A new, unreleased component called Information Matrix is being built in stages; slot components suggest a grid of customizable data cells, possibly for watch or widget surfaces.

  • Information Matrix: slot contracts addednew

    Add the Information Matrix slot component contracts

    The second step of the Information Matrix component, defining what each slot accepts before the visible pieces land.

  • Information Matrix: contracts addednew

    Add the Information Matrix contracts

    The first building block of the unreleased Information Matrix component; its purpose is not yet public.

  • Android XR: helper waits for the spatial scene

    Add awaitSpatialScene and embeddedWindowLeashes to SpatialSceneHelper

    SpatialSceneHelper gains awaitSpatialScene and embedded window leashes, groundwork for placing ordinary app windows inside XR scenes now that SceneCore is 1.0.

  • Compose: remember a style resolver

    Add `rememberStyleResolver()`

    A remembered style resolver fits the styles work above, where styles are resolved once and reused across recompositions.

  • People APIs hidden until API review

    Hide androidx.people APIs until API review

    A new androidx.people library is waiting for the API council; it matches the People Directory promised for Android 17 QPR2.

  • Glance Wear: fallback set trimmed to 1.6.0 operations

    Remove post-1.6.0 ops from Glance Wear SAFE_FALLBACK_SUPPORTED_OPERATIONS

    Wear widgets keep a safe list of operations older players understand; newer ones are removed so widgets degrade cleanly on old watches.

  • Lazy lists: slide animations for items

    Add Slide `EnterExitTransition` to `animateItem`

    animateItem gains slide enter and exit transitions, extending item animation beyond fade and placement.

  • Compose: an agent config for performance reviews

    Add Compose performance review agent config

    AndroidX checks in an agent configuration for reviewing Compose performance, mirroring the Compose performance skill announced at Dev Summit.

  • Embedded player: policy for unknown operations

    Add unknown-operation policy to embedded RcPlayer

    Compose Remote's embedded player decides what to do with operations it does not recognise, so documents from newer apps still render on older hosts.

  • Embedded player: optional features

    Add optional features to embedded RcPlayer

    The embedded player can declare optional capabilities, so a remote document can adapt to what the host supports.

  • AppState: DataStore-backed storage domain

    Add StorageDomain.DataStorePreference

    The new AppState library gets a storage domain backed by DataStore preferences, so state can persist across configurations and devices.

  • Glance adaptive: custom, excluded and action data

    Add CustomData, ExcludedData and action data

    Adaptive widget templates gain custom data, exclusions and action payloads, filling out the adaptive widget API before alpha.

  • Text: single-line line-height flag removed

    Remove isSingleLineLineHeightOptimizationEnabled flag

    The optimization behind the flag is now default, the kind of cleanup expected before Compose 1.13 stable.

  • Test Backup: restore, uninstall and device API ranges

    Add restoreOn, uninstallApp and @Device API ranges

    The backup testing library gains restoreOn, uninstallApp and per-API-range device rules, so backup and restore can be tested in CI.

  • Compose Remote: size to content in width and height

    Add RemoteIntrinsicSize and support in height/width modifiers

    Remote layouts get intrinsic sizing, so agent- or server-generated UI can wrap its content like local Compose.

  • CameraX: flash quirk for the Zebra TC77

    Add Zebra TC77 to ImageCaptureFlashNotFireQuirk

    A device quirk for a rugged scanner whose flash does not fire on capture; enterprise devices keep CameraX's quirk list growing.

Experts

17Read Kevin Schildhorn on hardening coroutine tests, skydoves' three-part series on how Compose compiles, remembers and recomposes, Vasiliy Zukanov on whether AI killed cross-platform after Shopify dropped React Native, and Joe Birch on morphing a button into a dialog in Compose Multiplatform.
Show 9 more

Library releases

10Metro 1.5.1 stops recompiling unrelated files and is a safe patch, OkHttp 5.5.1 and 4.12.1 are worth taking alongside this week's CVE mitigation write-up, Apollo Kotlin 4.4.4 fixes Java codegen for Long and Float scalars, kotlinx.serialization 1.12.0 and Ktor 3.6.1 are stable updates, and Decompose 3.6.0 is still in beta.
  • Compose Unstyled 2.10.2

    Patch Changes 649c471 : Make dialogs, modal bottom sheets, and modal drawers behave inside of a Host the same as without. Tab and Shift+Tab now stay inside the topmost modal, background controls receive no keystrokes, and closing the modal restores the previous focus. b59481b …

  • Metro 1.5.1new

    Fixes [IR/IC] Avoid recompiling unrelated files in the graph module when the constructor of an upstream @Contributes* class changes. [Gradle] Fix false hidden-dependency reports for contributions from associated compilations, such as main sources in tests. This particularly…

  • Metro 1.5.0

    New [metrox-viewmodel-compose] Add support for linuxX64 . [FIR] Add the diagnostic-level compiler option to change the severity of Metro diagnostics reported in FIR. It works like kotlinc's -Xwarning-level but can also lower errors and use IDE-only levels. See changing…

  • OkHttp 5.5.1

    Version 5.5.1

  • OkHttp 4.12.1

    Version 4.12.1

  • kotlinx.serialization 1.12.0

    This release is based on Kotlin 2.4.10 and contains all of the changes from 1.12.0-RC . Additionally, it includes a fix removing unnecessary leading whitespaces from Proguard rules ( #3270 ), and two changes stabilizing kotlin.uuid.Uuid serializer ( #3263 ) and…

  • Ktor 3.6.1new

    Release 3.6.1 ( #5970 )

  • Apollo Kotlin 4.4.4new

    A patch version to fix Java codegen generating bad code when using custom scalars mapped to Long or Float. 👷‍♂️ All changes [FIX] codegen: fix Java codegen generating bad code when using custom scalars mapped to Long or Float. ( #7032 )

Show 2 more
  • Decompose 3.6.0-beta02

    What's Changed Library Add linuxX64 and linuxArm64 targets by @arkivanov-bot in #1028 Fix wrong directions in AndroidPredictiveBackAnimatableV1 in RTL layout by @arkivanov in #1030 Fix wrong directions in AndroidPredictiveBackAnimatableV2 in RTL layout by @arkivanov in #1031 Add…

  • Roborazzi 1.77.0

    Bugfix: verifyRoborazzi* and compareRoborazzi* reuse the configuration cache With the configuration cache enabled, the second verifyRoborazzi* or compareRoborazzi* run after clean recalculated the task graph because "the file system entry 'build/outputs/roborazzi' has been…

Leaks & ecosystem

23Android 17 QPR3 Beta 3 brings back app lock with 26 Pixel fixes, Android 11 still runs on about one phone in ten, Samsung teases its Android XR glasses, and teardowns point to Android Auto remembering which number you call.
Show 15 more

Community

40A malformed SMS that crashes the phone service led Hacker News and Lobsters, JetBrains' Koog and a Claude Code app for Android trend in Kotlin, and Android Weekly covers OkHttp's CVE, Compose smoothness and shipping a Gemma model through Play.
Show 32 more
Start the Product & AI read4 stories, ranked by what you’d regret missing · 2 min read

  1. Mistral releases the weights of Mistral Large 4

    Another frontier-sized open model gives teams one more option to host AI themselves instead of renting a closed API, if they have the hardware.

    Mistral, the French AI lab, published the weights of Mistral Large 4, so anyone can download and run it. The Hugging Face listing is named one trillion, A 52 billion, which points to a mixture of experts design where only part of the model works on each word. It was the second biggest story on Hacker News, with about 1,260 points.

    HN front page · Across sourcesRead the sourceDiscussionProduct notes

  2. Impeccable, design rules for coding agents, adds nearly 3,000 stars

    Design taste is becoming a file you install, so a team can write its standards once and have every agent follow them.

    Impeccable is an open-source design vocabulary for coding agents. You steer an agent with commands like polish and bolder, and review its work with critique and audit, to catch generic AI-looking design before it ships. It gained about 2,900 stars in a day and is near 96 thousand, the fastest climber on GitHub today.

    GitHub Trending · Fastest risingRead the sourceProduct notes

  3. Hue lets teams test AI agents in clones of the apps they use

    As agents take real actions, a safe place to rehearse them is becoming its own product category.

    Hue, a new Y Combinator company, builds simulations for AI agents. These are stateful copies of the apps an agent works in, so you can watch what it does before it touches real accounts. It launched alongside Composal, which sends a swarm of agents to test a product like real users, and Simantic, which simulates firmware for engineers and agents.

    YC Launches · New ideaRead the sourceProduct notes

  4. A developer turned DOOM into a transformer model

    Games are turning into training data for models that simulate them, a cheap testbed for the world models robots and agents will need.

    A Show HN post describes using many DOOM games to create a transformer, the model design behind chatbots, instead of running DOOM on unusual hardware. On the same front page, a tiny world model, a model that generates a game as you play, ran on a Raspberry Pi Pico 2 W, and Product Hunt lists Genie 4, which turns a prompt into a world you can steer.

    Show HN · New ideaRead the sourceDiscussionProduct notes

Go deeper: themes, product notes, every source

Themes across sources

6Coding-agent control surfaces · Open-weight frontier models · Verifying agent work · Anti-slop design skills · On-device AI · World models

Ideas showing up in more than one place today. The line counts items per day over two weeks.

Coding-agent control surfaces

Rising · 11 items from 4 sources

GitHub is full of desktops for agents: CopilotKit's OpenBot gives each agent its own computer with every action logged, nightly-labs' openbot runs Codex, Claude and Grok side by side, OpenMausBot ships a VM for bots, and JetBrains' Thinkrail keeps climbing. Product Hunt adds MAO for putting several agents on one canvas, Yardlet and Mudroom for running them on a Mac or in a VM, Flatbread for sharing a whole agent run, and Arco for bringing GitHub and Linear issues into Codex. Mobile-Harness even puts Claude Code on an Android phone.

Why now. Teams now run several agents at once, so the open question has moved from which model to where agents live, what they can touch and how their work is reviewed.

CopilotKit/OpenBotGitHub Trending

More on this

Open-weight frontier models

Rising · 7 items from 2 sources

Mistral released the weights of Mistral Large 4, the second biggest story on Hacker News with about 1,260 points, and the model is already trending on Hugging Face. DeepSeek's V4.1 Flash stays the top trending model with community compressed builds, and MiniMax M3.1 Flash, Qwen3.8-27B and Nanbeige 4.3 fill out the chart.

Why now. Labs are competing on open releases at the top end, so self-hosting a frontier-sized model is now a choice about hardware, not access.

We're releasing the weights of Mistral Large 4HN front page

Verifying agent work

Rising · 9 items from 5 sources

Y Combinator's launches lean on testing agents before they act: Hue builds stateful clones of the apps an agent works in, Composal sends a swarm of agents to use a product like real users, and Simantic simulates firmware for engineers and agents. Product Hunt adds BotGauge and Dedica, GrayPass stops agents after login, a Dev.to post argues that human approval is not a security boundary unless it is tied to the exact action, and new papers benchmark CLI and computer-use agents against changing user needs.

Why now. Agents now take real actions in real systems, so rehearsing and checking them is becoming its own layer of tooling.

Hue: Simulations for AI agentsYC Launches

Anti-slop design skills

Rising · 7 items from 3 sources

Impeccable added about 2,900 stars today, the fastest climb on GitHub, with ai-design-skills close behind and Superpowers past 308 thousand stars. The skill format keeps spreading: Claude Code Game Studios packages 72 workflow skills, Product Hunt lists skills for building games and for designing App Store screenshots, and AndroidX checked in an agent config for Compose performance reviews.

Why now. Skills are a shared format across most coding agents, so one file of rules now reaches every agent a team uses.

pbakaus/impeccableGitHub Trending

More on this

On-device AI

Rising · 9 items from 4 sources

Liquid AI's LFM3, six 2B-class models for phones, laptops and robots, is on Product Hunt and trending on Hugging Face, NVIDIA's Nemotron 3.5 Bolt fits in 4 GB, and Hob pitches an open-source, on-device alternative to Claude and ChatGPT. On Android, LiteRT in Google Play services offers faster GPU inference, and Android Weekly covers Gemini Nano's input limit, shipping a 584 MB Gemma model through Play and on-device agents with Google's ADK.

Why now. Small open models now handle everyday tasks, and the runtimes to serve them are shipping inside the platform itself.

LFM3 by Liquid AIProduct Hunt

World models

Steady · 4 items from 3 sources

A Show HN post turned many DOOM games into a transformer, a tiny world model ran games on a Raspberry Pi Pico 2 W on the Hacker News front page, and Product Hunt lists Genie 4, which turns a prompt into a world you can steer, and Viggle JST-2 for 3D scenes and motion from one model.

Why now. Games give unlimited, labelled training data, so models that simulate an environment frame by frame have become a hobby-scale project as well as a lab one.

Show HN: I Turned DOOM into a TransformerShow HN

Fading: Agent-ready Android (last seen 2026-10-29) · Android XR glasses (last seen 2026-10-29) · Agent supply-chain safety (last seen 2026-10-27) · Agents as security auditors (last seen 2026-10-26) · Open voice models (last seen 2026-10-26) · Agent sandboxes (last seen 2026-10-25)

Product notes

16Who each product is for, the problem, why now, its wedge and how it makes money. Today’s picks first.

Plain text: the source says soDotted: inferred

We're releasing the weights of Mistral Large 4

discussionHN front page1,262 pts · 355 comments

What
Mistral released the weights of Mistral Large 4. The Hugging Face listing is named 1T-A52B, which points to about one trillion total and 52 billion active parameters in a mixture-of-experts design.
For whom
Companies that need a large model on their own infrastructure for cost, privacy or regulation reasons.
Problem
Frontier-quality models have mostly been API-only, which rules them out for data that cannot leave a company.
Why now
Mixture-of-experts designs make very large models cheap enough to serve, and open releases from several labs raised the bar.
Wedge
Open weights at the top end, for European and regulated buyers.
Money
Paid API, enterprise deployments and support on top of open weights.
Like
DeepSeek V4.1 · Qwen3.8 · MiniMax M3.1 · Llama

LessonGiving away the model can sell the service around it; the moat moves to deployment and trust.

pbakaus/impeccable

repoGitHub Trending★ 2,870 today

What
Impeccable, a design vocabulary for coding agents: steer with commands like polish and bolder, review with critique and audit, and catch generic AI-looking design before it ships. About 2,900 stars today, near 96 thousand total.
For whom
Developers building front ends with agents who lack a designer on every change.
Problem
Agent-built interfaces look alike; teams fix them by hand or accept the look.
Why now
Skills became a portable format across agents, so design taste can ship as a file.
Wedge
A small set of named commands that are easy to remember and repeat.
Money
Open source, no model yet.
Like
ai-design-skills · Hallmark · Superpowers

LessonGive users verbs, not settings: a named command is easier to adopt than a style guide.

Hue: Simulations for AI agents

productYC Launches63 votes

What
Hue builds simulations for AI agents: stateful clones of the apps an agent works in, so you can test it before it touches real systems.
For whom
Teams shipping agents that act in business software such as CRMs, ticketing or payments.
Problem
Testing an agent against real accounts is risky and slow, and mocks miss how state changes over a task.
Why now
Agents now run multi-step tasks in real tools, so a mistake costs more than a bad answer.
Wedge
Clones of the specific apps an agent touches, not a general sandbox.
Money
Likely usage-based pricing per simulation or seat.
Like
Composal · Simantic · BotGauge · staging environments

LessonA new kind of worker needs a new kind of test environment; pick one environment and make it feel real.

Show HN: I Turned DOOM into a Transformer

productShow HN, HN front page57 pts · 13 comments

What
A Show HN project that uses many DOOM games to create a transformer, instead of running DOOM on unusual hardware.
For whom
Developers curious about how world models are trained, with a familiar game as the example.
Problem
World-model research is mostly closed and expensive, so few people see how one is built.
Why now
Consumer GPUs and open training code make small game models a weekend-scale project.
Wedge
A meme everyone knows, DOOM runs on everything, flipped into a model.
Money
Hobby project, no model.
Like
GameNGen · Genie 4 · tiny world model on Pico 2 W

LessonA familiar frame turns a hard topic into one people click and learn from.

Grab Reduces ANRs by 32% and improves startup time by 35% with R8 and Profile guided optimization

articleAndroid Developers Blog, Android WeeklyAndroid

What
A Google case study: Grab, the Southeast Asian superapp for deliveries, rides and payments, cut ANRs by 32 percent and improved startup time by 35 percent using R8 and profile-guided optimization.
For whom
Teams with a large, multi-feature Android app where startup and ANRs are tracked as business metrics.
Problem
Big apps grow slow and freeze-prone as features pile up; teams cope with manual lazy loading and one-off fixes.
Why now
R8 optimization and baseline profiles are now default-friendly in the toolchain, so the remaining work is configuration and keep rules, not custom engineering.
Wedge
Build-time changes that need no feature rewrites.
Money
Free platform tooling; Google benefits when Android apps feel faster.
Like
Baseline Profiles · Startup Profiles · Macrobenchmark · DexGuard

LessonThe cheapest performance win is often a build setting your team never turned on; measure before and after so the number sells the change.

JetBrains/koog

repoGitHub Trending · KotlinAndroid★ 5 today

What
Koog, JetBrains' framework for building AI agents in Java and Kotlin that are predictable and fault-tolerant, running anywhere from backend services to Android.
For whom
Kotlin teams that want agent logic in the same language and codebase as their app or server.
Problem
Most agent frameworks are Python or TypeScript, so Kotlin teams wrap HTTP calls by hand or run a separate service.
Why now
Agents moved from demos into production apps, where retries, typed tools and observability matter more than prompts.
Wedge
Kotlin Multiplatform reach, one agent definition from server to phone.
Money
Open source; JetBrains benefits through its IDEs, Junie and AI subscriptions.
Like
LangChain4j · Spring AI · Google ADK · LangGraph

LessonOwning the language gives you a natural place to own the framework; meet developers where their code already lives.

techjarves/Mobile-Harness

repoGitHub Trending · KotlinAndroid★ 8 today

What
Mobile-Harness runs Claude Code on Android: chat with a coding agent, run Linux commands, edit files, review diffs and preview web apps on the phone.
For whom
Developers who want to check on or steer coding work away from a laptop.
Problem
Agent work runs for minutes or hours, and today people SSH from a phone terminal or wait until they are back at a desk.
Why now
Agents do long tasks on their own, so the human mostly reviews and approves, which fits a phone screen.
Wedge
Diff review and previews on the phone rather than a full IDE.
Money
Open source, no model yet.
Like
Termux · AfterLid · Claude mobile app · GitHub Mobile

LessonWhen the work shifts from typing to reviewing, the device that wins changes too.

CopilotKit/OpenBot

repoGitHub Trending★ 247 today

What
CopilotKit's OpenBot: open-source AI coworkers that each get their own computer, a browser, files and tools, with every action decided before it happens and recorded after; works with any model.
For whom
Teams that want agents doing ongoing work but need an audit trail of what each one did.
Problem
Agents sharing a user's machine and accounts are hard to isolate and harder to audit afterwards.
Why now
Computer-use models are good enough to work in a browser, so isolation and logging became the blocker.
Wedge
A recorded decision log per action, aimed at trust rather than raw capability.
Money
Open source; CopilotKit likely sells hosted or enterprise versions.
Like
nightly-labs openbot · OpenMausBot · Paddock · Mudroom

LessonIn a crowded agent market, the audit log can be the product; sell the part that makes buyers comfortable saying yes.

Composal: A swarm of agents that tests your product like real users

productYC Launches6 votesHype flag

What
Composal sends a swarm of agents to test your product like real users, to catch bugs before your users do.
For whom
Product teams without a large QA team who ship web or app changes often.
Problem
Scripted end-to-end tests only cover the paths someone wrote down; real users find the rest.
Why now
Computer-use agents can now operate a product's UI without scripts.
Wedge
Exploratory testing, the part scripted tests do not cover.
Money
Likely a subscription per product or test run.
Like
Maestro · QA Wolf · Hue · Mabl

Like real users is a broad claim; how bugs are judged and how many are false alarms is not shown.

LessonSell the result buyers already pay for, bugs found before release, not the agent behind it.

Launch HN: Simantic (YC F26) – Firmware Simulation for AI Agents

discussionHN front page56 pts · 16 comments

What
Simantic, a Y Combinator company, is building a simulation layer that lets engineers and AI agents test and validate firmware.
For whom
Embedded and hardware teams whose firmware testing waits on physical boards.
Problem
Firmware tests need real hardware on a bench, which is slow, scarce and hard to automate.
Why now
Coding agents now write firmware too, and they need a fast loop to check it without a board.
Wedge
Simulation built for agents as well as people.
Money
Likely a subscription for simulation capacity.
Like
Renode · QEMU · Wokwi

LessonWhen a new user type, agents, arrives in an old field, rebuilding the tools for them is a startup.

Show HN: Auto Dismiss Mac OS "Allow <app> to find devices on local networks?"

productShow HN, HN front page113 pts · 85 comments

What
A small macOS tool that automatically dismisses the Allow app to find devices on local networks popup, which has no global off switch.
For whom
Mac users who keep seeing the local network prompt and almost never want to allow it.
Problem
There is no system setting to turn the prompt off, so people click Don't Allow over and over.
Why now
More apps probe the local network, so the prompt now appears often enough to annoy.
Wedge
One narrow annoyance with a one-line pitch.
Money
Free tool, no model.
Like
Little Snitch · LuLu

LessonA permission prompt users always answer the same way is a design failure; on Android too, ask only when the answer is not obvious.

Show HN: Lasm – Write formally verified webapps in Lean for Node/serverless/edge

productShow HN, HN front page15 pts · 0 comments

What
Lasm lets you write formally verified web apps in Lean and run them on Node, serverless or edge platforms.
For whom
Developers who want proofs that their backend logic is correct, not just tests.
Problem
Formal verification lives in research tools far from the runtimes web apps actually ship on.
Why now
The author argues formal verification is the next step for software; agents writing code make machine-checked correctness more valuable.
Wedge
Deploy targets developers already use, not a new runtime.
Money
Open source, no model yet.
Like
F* · Dafny · Lean 4

LessonTo spread a rigorous tool, ship it where people already deploy.

LFM3 by Liquid AI

productProduct Hunt#43 on Product Hunt

What
LFM3 by Liquid AI: six 2B-class models for phones, laptops and robots, with LFM3-1.7B and LFM3-Omni-2B trending on Hugging Face.
For whom
App and device makers who want language and multimodal features without a server round trip.
Problem
Cloud inference adds cost, latency and privacy concerns for features that run on every tap.
Why now
Small models now handle everyday tasks, and phone NPUs and GPU runtimes can serve them.
Wedge
A family sized for specific devices rather than one general model.
Money
Open weights with commercial licensing and enterprise deals.
Like
Gemma · Nemotron 3.5 Bolt · Qwen3.5 small · Gemini Nano

LessonSizing a product to the hardware your customer already owns removes the biggest objection before it is raised.

HKUDS/DeepTutor

repoGitHub Trending★ 1,930 today

What
DeepTutor, an open-source project for lifelong personalized tutoring, the top GitHub repo today with about 1,900 new stars and about 50 thousand total.
For whom
Learners and education builders who want a tutor that remembers a student over time.
Problem
Chatbots forget a student between sessions, so every lesson starts from zero.
Why now
Agent memory layers make a tutor that keeps a long-term picture of a learner practical.
Wedge
Lifelong personalization as the headline, not answering questions.
Money
Open source with a hosted site; no model stated.
Like
Khanmigo · Google LearnLM · Duolingo Max

LessonIn education, memory of the learner is the moat; the model is replaceable.

Flatbread

productProduct Hunt#3 on Product Hunt

What
Flatbread lets you share a whole agent run, not a screenshot.
For whom
Developers and teams who need to show a colleague what an agent did and why.
Problem
People paste screenshots or long logs, which lose the steps and context of a run.
Why now
Agent runs are now long and important enough to review, debug and teach from.
Wedge
A share link, the simplest possible collaboration feature.
Money
Likely free sharing with paid team features.
Like
Loom · LangSmith traces · GitHub Gists

LessonEvery new medium needs its own share link; the first one people use becomes the default.

StoreLocale

productProduct Hunt#25 on Product Hunt

What
StoreLocale translates the app interface shown in App Store and Google Play screenshots.
For whom
Indie and small app teams that localize their app but not their store listing images.
Problem
Localized screenshots mean re-capturing or re-editing every image per language, so most teams ship English screenshots everywhere.
Why now
Image models can now edit text inside a screenshot cleanly.
Wedge
One tedious store-listing chore rather than full localization.
Money
Likely pay per app or per language.
Like
App Store Screenshots · Fastlane screengrab · AppLaunchpad

LessonFind the step everyone skips because it is tedious; that skipped step is a product.

By source

11Everything collected today from GitHub Trending, Product Hunt, Show HN, HN front page, YC Launches, HF trending models, HF trending spaces, HF daily papers, Lobsters, Dev.to top, Techmeme, in each source’s own order.

GitHub Trending

11
  • HKUDS/DeepTutorcard

    DeepTutor: Lifelong Personalized Tutoring. https://deeptutor.info/ .

    ★ 1,930 today

  • obra/superpowers

    Superpowers, the agentic skills framework, passes 308 thousand stars.

    ★ 1,034 today

  • pbakaus/impeccablecard

    The design vocabulary for your coding agent. Steer with polish and bolder, review with critique and audit, catch AI slop before it ships.

    ★ 2,870 today

  • JetBrains/thinkrail

    JetBrains' Thinkrail, a lightweight IDE built around a coding agent, adds about 1,100 stars.

    ★ 1,117 today

  • milind-soni/OpenMausBotnew

    OpenMausBot, an open-source bot with its own virtual machine, adds about 1,200 stars.

    ★ 1,199 today

  • nightly-labs/openbot

    openbot runs Codex, Claude and Grok as persistent teammates with task queues and browser control on your desktop.

    ★ 709 today

  • elayadesign/ai-design-skills

    ai-design-skills adds about 1,200 stars, more design taste packaged for agents.

    ★ 1,187 today

  • thedotmack/claude-memnew

    claude-mem compresses what an agent did and feeds it back next session, past 100 thousand stars.

    ★ 336 today

Show 3 more
  • CopilotKit/OpenBotnewcard

    Open-source AI coworkers that each get a computer of their own: a browser, files and tools, with every action decided before it happens and recorded after. Bring any AG-UI agent.

    ★ 247 today

  • togg53192-cmd/jailbreaks

    LIST OF ALL MY JAILBREAKS

    ★ 1,594 today

  • Donchitos/Claude-Code-Game-Studiosnew

    Claude Code Game Studios packages 49 agents and 72 skills as a game studio hierarchy.

    ★ 353 today

Product Hunt

49
  • Notch Metricsnew

    Build analytics widgets for the Mac notch with Claude Code, the top Product Hunt launch.

    #1 on Product Hunt

  • Polylane for ChatGPTnew

    Ask ChatGPT what broke in production, incident data brought into the chat app.

    #2 on Product Hunt

  • Flatbreadnewcard

    Share the whole agent run, not a screenshot

    #3 on Product Hunt

  • Scapenew

    An email inbox that sorts and acts on mail for you.

    #4 on Product Hunt

  • Radarnew

    Radar gives your existing agent production context instead of selling a separate AI SRE.

    #5 on Product Hunt

  • Kinnew

    Kin pitches user-owned memory so AI labs answer to the user.

    #6 on Product Hunt

  • Reset Alertsnew

    Text alerts when your Codex usage resets, so paid quota is not wasted.

    #7 on Product Hunt

  • Smart Jellynew

    Everyday AI tools and memory in one Mac app.

    #8 on Product Hunt

Show 41 more
  • Arconew

    GitHub and Linear issues inside Codex, so the agent sees the ticket.

    #9 on Product Hunt

  • Floot in ChatGPTnew

    Build and publish apps without leaving ChatGPT, apps inside the chat app.

    #10 on Product Hunt

  • Gologin Cloud Browsernew

    A cloud browser for agents that need to get through difficult sites.

    #11 on Product Hunt

  • TopAgentnew

    An agent that drives attendance for B2B events.

    #12 on Product Hunt

  • OutreachGuynew

    An AI sales rep that follows up on leads you dropped.

    #13 on Product Hunt

  • Suno MCPnew

    Suno's music generation exposed as an MCP server for agents.

    #14 on Product Hunt

  • Voxel51 Multimodalnew

    Voxel51 extends its data tooling to multimodal robot data.

    #15 on Product Hunt

  • Viggle JST-2new

    Viggle JST-2 makes 3D scenes, characters and motion from one model.

    #16 on Product Hunt

  • ChatInboxnew

    An inbox grouped by contact instead of by thread.

    #17 on Product Hunt

  • Enclavenew

    Enclave puts the context you need behind a single hover.

    #18 on Product Hunt

  • Intrascope 2.0new

    Intrascope 2.0 adds an agent, 1,650 app connections and model routing.

    #19 on Product Hunt

  • MAOnew

    MAO puts Claude Code, Codex and Grok on one shared canvas.

    #20 on Product Hunt

  • AfterLidnew

    Keeps a Mac working with the lid closed and lets you check in from a phone, built for long agent runs.

    #21 on Product Hunt

  • IFTTT.ainew

    IFTTT relaunches around AI automation.

    #22 on Product Hunt

  • Confettinew

    Confetti designs hardware gadgets from plain English, also a YC launch.

    #23 on Product Hunt

  • AutoShelf 2.6new

    AutoShelf 2.6 automates Mac folders with if-this-then-that rules.

    #24 on Product Hunt

  • StoreLocalenewcard

    Translate app UI in App Store & Google Play screenshots

    #25 on Product Hunt

  • AssemblyAInew

    AssemblyAI adds text-to-speech built for voice agents.

    #26 on Product Hunt

  • Empacernew

    A voice-first assistant aimed at small business owners.

    #27 on Product Hunt

  • JIMMYnew

    JIMMY, an agent that runs on your PC.

    #28 on Product Hunt

  • Recall for Macnew

    Recall for Mac records what happens on your Mac so you can search it later.

    #29 on Product Hunt

  • Gemini 4 Carbonnew

    Gemini 4 Carbon, Google's latest frontier model, on Product Hunt.

    #30 on Product Hunt

  • Trigger.dev: Webhooks for your AI agents

    Trigger.dev adds webhooks so agents react to Slack, Supabase and 50 other services.

    #31 on Product Hunt

  • Hob

    Hob, an open-source on-device alternative to Claude and ChatGPT.

    #32 on Product Hunt

  • OpenLLM

    OpenLLM promises to take your AI tokens to any provider.

    #33 on Product Hunt

  • Congress Trades

    Shows how members of Congress trade and whether copying them pays.

    #34 on Product Hunt

  • Our Silly Games

    Skills for building games with Claude plus free multiplayer hosting.

    #35 on Product Hunt

  • NotiKeep

    Notification history for iPhone kept on device, a feature Android has had for years.

    #36 on Product Hunt

  • App Store Screenshots

    Your coding agent designs your App Store screenshots.

    #37 on Product Hunt

  • Tulu

    Autonomous marketing operations and analytics for small businesses.

    #38 on Product Hunt

  • Open For sales

    A sales campaign from one prompt.

    #39 on Product Hunt

  • Genie 4

    Genie 4 turns a prompt into a world you can steer.

    #40 on Product Hunt

  • bytemail

    A native Mac email client in the notch that strips trackers.

    #41 on Product Hunt

  • RageLayer

    Lets visitors smash your website and sweeps it back, a marketing toy.

    #42 on Product Hunt

  • LFM3 by Liquid AIcard

    Six 2B-class models for phones, laptops and robots

    #43 on Product Hunt

  • Mudroom

    Mudroom runs coding agents in a VM and lets you review changes like a pull request.

    #44 on Product Hunt

  • BotGauge

    BotGauge calls itself a trust layer for agents.

    #45 on Product Hunt

  • Yardlet

    Yardlet, a Mac app for the coding agents you already use.

    #46 on Product Hunt

  • Dedica

    Dedica checks whether agents can use your website, accessibility for a new kind of visitor.

    #47 on Product Hunt

  • Basedash View-only members

    Basedash adds view-only members for sharing dashboards.

    #48 on Product Hunt

  • TONKO

    A pocket money app that teaches kids about money.

    #50 on Product Hunt

Show HN

9
Show 1 more

HN front page

30
Show 22 more

YC Launches

20
Show 12 more

HF trending models

19
Show 11 more

HF trending spaces

15
Show 7 more

HF daily papers

20
Show 12 more

Lobsters

25
Show 17 more

Dev.to top

20
Show 12 more

Techmeme

15
Show 7 more

About this issue

403 items from 77 sources, 7 of them seen in several sources. 3 sources failed to fetch.

Source health

77Which sources were fetched today, and how many items each returned.
SourceSectionItemsStatus
GitHub TrendingMain11Fetched
Product HuntMain50Fetched
Show HNMain9Fetched
HN front pageMain30Fetched
YC LaunchesMain20Fetched
HF trending modelsMain20Fetched
HF trending spacesMain20Fetched
HF daily papersMain20Fetched
LobstersMain25Fetched
Dev.to topMain20Fetched
TechmemeMain15Fetched
AndroidX releasesAndroid · Official & releases58Fetched
Android Developers BlogAndroid · Official & releases5Fetched
Android Studio releasesAndroid · Official & releases5Fetched
Kotlin blogAndroid · Official & releases0Fetched
Kotlin releasesAndroid · Official & releases0Fetched
Compose Multiplatform releasesAndroid · Official & releases1Fetched
Android Developers YouTubeAndroid · Official & releases0Failed
Kotlin on YouTube (JetBrains)Android · Official & releases0Failed
Google Developers BlogAndroid · Official & releases1Fetched
AndroidX merged API changesAndroid · AOSP / AndroidX insider27Fetched
Dove Letter (skydoves)Android · Experts5Fetched
skydoves on MediumAndroid · Experts0Fetched
Jake WhartonAndroid · Experts0Fetched
Zac SweersAndroid · Experts0Fetched
Chris BanesAndroid · Experts2Fetched
Ryan HarterAndroid · Experts0Fetched
Talking KotlinAndroid · Experts0Fetched
Joe BirchAndroid · Experts2Fetched
CommonsWare (Mark Murphy)Android · Experts3Fetched
TechYourChance (Vasiliy Zukanov)Android · Experts1Fetched
Kt. Academy (Marcin Moskała)Android · Experts2Fetched
Touchlab (Kotlin Multiplatform)Android · Experts2Fetched
Landscapist releasesAndroid · Library releases0Fetched
Balloon releasesAndroid · Library releases0Fetched
Sandwich releasesAndroid · Library releases0Fetched
Cloudy releasesAndroid · Library releases0Fetched
Compose Stability Analyzer releasesAndroid · Library releases0Fetched
Haze releasesAndroid · Library releases0Fetched
Compose Unstyled releasesAndroid · Library releases1Fetched
Coil releasesAndroid · Library releases0Fetched
Circuit releasesAndroid · Library releases0Fetched
Metro releasesAndroid · Library releases2Fetched
Molecule releasesAndroid · Library releases0Fetched
Turbine releasesAndroid · Library releases0Fetched
Paparazzi releasesAndroid · Library releases0Fetched
SQLDelight releasesAndroid · Library releases0Fetched
OkHttp releasesAndroid · Library releases2Fetched
Retrofit releasesAndroid · Library releases0Fetched
LeakCanary releasesAndroid · Library releases0Fetched
Koin releasesAndroid · Library releases0Fetched
Dagger / Hilt releasesAndroid · Library releases0Fetched
KSP releasesAndroid · Library releases0Fetched
kotlinx.coroutines releasesAndroid · Library releases0Fetched
kotlinx.serialization releasesAndroid · Library releases1Fetched
Ktor releasesAndroid · Library releases1Fetched
Apollo Kotlin releasesAndroid · Library releases1Fetched
Kermit releasesAndroid · Library releases0Fetched
Store releasesAndroid · Library releases0Fetched
Decompose releasesAndroid · Library releases1Fetched
Orbit MVI releasesAndroid · Library releases0Fetched
Roborazzi releasesAndroid · Library releases1Fetched
detekt releasesAndroid · Library releases0Fetched
ktlint releasesAndroid · Library releases0Fetched
MockK releasesAndroid · Library releases0Fetched
Lottie releasesAndroid · Library releases0Fetched
Authority Insights (leaks & teardowns)Android · Leaks & ecosystem8Fetched
9to5Google · AndroidAndroid · Leaks & ecosystem8Fetched
Android AuthorityAndroid · Leaks & ecosystem6Fetched
Android FaithfulAndroid · Leaks & ecosystem1Fetched
Android WeeklyAndroid · Community15Fetched
GitHub Trending · KotlinAndroid · Community8Fetched
Lobsters · androidAndroid · Community1Fetched
HN · AndroidAndroid · Community14Fetched
ProAndroidDevAndroid · Community5Fetched
Kotlin WeeklyAndroid · Community1Fetched
Philipp Lackner (YouTube)Android · Community0Failed
1 of 5

Grab cuts ANRs by a third with R8 and startup profiles

If your app ships without R8 optimization or baseline profiles, this is one of the cheapest speed and stability wins left on Android.

Google's Android Developers Blog published a case study on Grab, the Southeast Asian superapp for rides, food delivery and payments. Grab used R8, the tool that shrinks and optimizes app code, together with profile guided optimization, which prepares the code users run most. The result was about 32 percent fewer ANRs, the app not responding errors, and about 35 percent faster startup.

Android Developers Blog · OfficialRead the source

1 of 4

Mistral releases the weights of Mistral Large 4

Another frontier-sized open model gives teams one more option to host AI themselves instead of renting a closed API, if they have the hardware.

Mistral, the French AI lab, published the weights of Mistral Large 4, so anyone can download and run it. The Hugging Face listing is named one trillion, A 52 billion, which points to a mixture of experts design where only part of the model works on each word. It was the second biggest story on Hacker News, with about 1,260 points.

HN front page · Across sourcesRead the sourceDiscussion