--- title: "Expo SDK 57, Hermes V1+, agent-device Replay, GPT-5.6, Grok 4.5 RN Evals, Apex launch, Octane" description: "Cloud device testing, smarter agent replays, React Navigation 8 progress, and more CSS in React Native core." url: https://www.callstack.com/newsletters/expo-sdk-57-hermes-agent-device-replay-apex-and-react-native-evals updated: 2026-10-24T07:18:04.735Z --- # Expo SDK 57, Hermes V1+, agent-device Replay, GPT-5.6, Grok 4.5 RN Evals, Apex launch, Octane Cloud device testing, smarter agent replays, React Navigation 8 progress, and more CSS in React Native core. _By Michał Pierzchala, Principal Engineer @ Callstack · Published 2026-08-05_ Hey, It's Michał with July's edition of the Callstack newsletter. Expo SDK 57 is here, CSS gradients are coming to React Native 0.87, and Hermes has a new stable release you can try today if you're feeling brave. Meanwhile, the Incubator team has been busy with agent-device replays, cloud device testing, fresh React Native Evals, Apex improvements, and more work landing in React Native core. It's a packed one. Let's go! ## Community highlights - [**Expo SDK 57**](https://expo.dev/changelog/sdk-57): Expo is now tracking React Native releases more closely and skipping the beta phase for SDK 57. The release includes React Native 0.86, Android edge-to-edge improvements, light and dark mode emulation in React Native DevTools, plus a healthy batch of rendering, layout, and animation fixes. Bold. - [**A safer SafeAreaProvider moves toward React Native core**](https://github.com/react/react-native/pull/57500): The proposed provider uses native safe-area information with synchronous initialization. It aims to reduce first-frame flicker and remove the need to set up the provider manually. I thought we’re making the core leaner though. - [**Background gradients are coming to React Native 0.87**](https://github.com/react/react-native-website/pull/5191): The `backgroundImage` style is becoming stable, bringing linear and radial gradients to React Native without another dependency. - [**Hermes ships a new stable release**](https://github.com/facebook/hermes/blob/static_h/doc/blog/2026-06-05-new-hermes-stable-release.md): The `260318099.0.1` release brings major JSON performance gains, Set methods, Iterator Helpers, and TypeScript transform support. New Hermes versions are [not immediately available in React Native](https://reactnative.dev/releases/branches), but Tzvetan from the core team recently [published a guide](https://github.com/tmikov/rn-0.xx-hermes-v1/tree/master/targets/rn-0.86) for testing it in your app if you're brave enough. - [**Octane compiles React without the virtual DOM**](https://octanejs.dev/): Octane keeps the familiar JSX and Hooks programming model, then compiles it ahead of time into direct DOM operations. I'm curious how far it can push React's programming model with less runtime work. - [**npm adds publish-time malware scanning**](https://github.blog/changelog/2026-07-28-npm-publish-time-malware-scanning-and-dual-use-metadata/): npm now scans packages before they become available on the registry. Publishing may take a few extra minutes, which feels like a fair trade for a healthier ecosystem. ## Callstack Incubator highlights ### agent-device July was a busy month for agent-device. [v0.20 introduced replay](https://x.com/thymikee/status/2080297621491483014), which turns a journey discovered by an agent into a deterministic script that can run again in seconds without another model call. Replays have existed since the early agent-device releases, but we found a better way to use them: hybrid workflows. The agent runs a script through the predictable parts and takes over when the flow needs intelligence. Think E2E tests, but smart. > > Agents can now replay what they already know, and think only when they need to. > > > > • 65s → 6s on the same flow > > • faster feedback loops > > • fewer tokens > > > > Give your agents better tools: > > > > » agent-device session save-script [pic.twitter.com/lmXjZTltyl](https://t.co/lmXjZTltyl) > > > > — Michał Pierzchała (@thymikee) [July 25, 2026](https://x.com/thymikee/status/2081006447626760512?ref%5Fsrc=twsrc%5Etfw) [Post on X](https://x.com/thymikee/status/2081006447626760512) We followed that with [v0.20.1](https://x.com/thymikee/status/2081792001301754073), which improved stability, added VegaOS support, expanded physical iOS coverage, and made secrets safer during recording and replay. Your agent can test iOS Widgets and Live Activities now, too. The new `--settle` flag waits for a usable UI state and returns the resulting observation in one step. That means fewer tool calls and tokens, with higher success rates for models such as Haiku. I'm especially excited about direct support for remote device clouds. With a BrowserStack, AWS Device Farm, or Limrun subscription, an agent running inside a cloud container, [including Devin](https://x.com/thymikee/status/2072709478634459531), can access a real device for verification. Fully cloud-based mobile development is here. > > AI agents can now connect to real cloud devices🔥 > > > > agent-device now supports BrowserStack and AWS Device Farm. > > > > Test, inspect, and debug on real devices in cloud instead of stopping at local simulators. > > > > » agent-device connect browserstack > > » agent-device connect aws-device-farm [pic.twitter.com/A0kR7aVJEI](https://t.co/A0kR7aVJEI) > > > > — Michał Pierzchała (@thymikee) [July 1, 2026](https://x.com/thymikee/status/2072338300744945878?ref%5Fsrc=twsrc%5Etfw) [Post on X](https://x.com/thymikee/status/2072338300744945878) agent-device is also featured in the [Expo documentation](https://docs.expo.dev/agents/agent-device/) as a recommended tool for agents developing and verifying Expo apps. We also played with Fable 5, [which made relaunching iOS apps six times faster](https://x.com/thymikee/status/2072767578640933318). Agents relaunch apps a lot. It's a good model. ### React Native Evals [New React Native Evals results are live](https://x.com/thymikee/status/2077456508728627302), with five fresh models: GPT 5.6 Sol, Terra, Luna, Cursor Grok 4.5, and Meta's Muse Spark 1.1. Under the hood, an automated Airflow and MLflow pipeline now produces the artifacts and analyzer outputs. Our goal is to publish benchmarks within 24 hours of a model becoming available through AI Gateway. > > React Native Evals are out with a bunch of new models: > > > > → Sol only slightly better at 3x cost compared to Terra or Grok 4.5\. > > → Muse Spark 1.1 feels like OpenAI's fourth model in terms of perf vs cost > > → Grok 4.5 scores 12 pp higher than Grok 4 at 4x less cost [pic.twitter.com/19BSvMuIgu](https://t.co/19BSvMuIgu) > > > > — Michał Pierzchała (@thymikee) [July 15, 2026](https://x.com/thymikee/status/2077456508728627302?ref%5Fsrc=twsrc%5Etfw) [Post on X](https://x.com/thymikee/status/2077456508728627302) The research side reached a bigger milestone: our [AAAI-27 submission](https://openreview.net/forum?id=SNV3zvpsxa) standardizes 66 tasks and 273 requirements across five categories, with ten runs per system and 22 comparable systems. We're also building a new set of evals for Expo SDK and Expo Router. > > Working on a new set of evals for our React Native Evals. Will be covering parts of Expo SDK, most common Expo Router flows and Expo UI. Results should be landing next week 🔜 [pic.twitter.com/DTzKU6ms1u](https://t.co/DTzKU6ms1u) > > > > — Mike (@grabbou) [July 3, 2026](https://x.com/grabbou/status/2072963578869784675?ref%5Fsrc=twsrc%5Etfw) [Post on X](https://x.com/grabbou/status/2072963578869784675) ### Apex Apex runs on Gemma 4. We recently tested and improved Gemma's tool calling and chat template, with [our changes flowing back upstream](https://x.com/thymikee/status/2077481876738982154). We plan to open Apex to everyone on the waitlist this week, with GA following later in August. Get on the waitlist if you haven't already. > > Thanks for the updates, and love that our feedback made it in! Now it's [@callstackio](https://x.com/callstackio?ref%5Fsrc=twsrc%5Etfw)'s turn to cook - we're bringing these improvements into Apex, our React Native model built on Gemma 4: > > > > — Piotr Miłkowski (@xeonbloomfield) [July 15, 2026](https://x.com/xeonbloomfield/status/2077478725705212355?ref%5Fsrc=twsrc%5Etfw) [Post on X](https://x.com/xeonbloomfield/status/2077478725705212355) Piotr is also experimenting with a pipeline that keeps React Native and mobile development knowledge current without another round of fine-tuning. Expect a blog post on that later in August. > > Apex by [@callstackio](https://x.com/callstackio?ref%5Fsrc=twsrc%5Etfw) just got an update. > > > > The latest RN libraries? ✅ > > Runs on this week's Gemma update? ✅ > > Fixes for reported issues? ✅ > > > > Wide beta starts next week! > > > > Request access here: [pic.twitter.com/FxoyjH4Zm7](https://t.co/FxoyjH4Zm7) > > > > — Piotr Miłkowski (@xeonbloomfield) [July 17, 2026](https://x.com/xeonbloomfield/status/2078139693493895607?ref%5Fsrc=twsrc%5Etfw) [Post on X](https://x.com/xeonbloomfield/status/2078139693493895607) ### React Navigation 8 React Navigation 8 keeps taking shape. July brought [another progress update](https://x.com/reactnavigation/status/2074820941444509824), SSR and hydration fixes, and [a working Vite SSR example](https://x.com/satya164/status/2079231984707019247). We also refactored the native package so web projects no longer need React Native Web just to use it, and improved concurrency in `useNavigationState`. > > 🚀 React Navigation 8 keeps leveling up! > > > > Our latest blog post rounds up everything shipped since March: > > > > ⚛️ Suspense & concurrent rendering > > 📡 Experimental data loaders > > 🔠 Typed hooks in dynamic navigators > > 🔗 Shared paths for deep links > > 📌 Retaining screens in Stack > > 🎨 Material… > > > > — React Navigation (@reactnavigation) [July 8, 2026](https://x.com/reactnavigation/status/2074820941444509824?ref%5Fsrc=twsrc%5Etfw) [Post on X](https://x.com/reactnavigation/status/2074820941444509824) ### Brownfield and migration tooling React Native Brownfield added Expo SDK 56 support, laid the groundwork for Android Gradle Plugin 9, fixed duplicate native libraries during cold builds, and gained scheduled testing against Expo previews. For reasons known to the team (and now to me), Brownfield also supports MAUI. You can embed React Native views inside a MAUI app, which makes incremental migrations to React Native possible. > > Part 3: MAUI(xamarin) 🤝 React Native > > > > Brownfield integrations always require sharing data like user profiles and tokens. But syncing state across frameworks can quickly turn into a nightmare.💥 > > > > Watch @callstack/brownie handle it seamlessly using a folly c++ store 🎥👇 [pic.twitter.com/X5VJlGvGTd](https://t.co/X5VJlGvGTd) > > > > — Hur Ali (@hurali97) [July 10, 2026](https://x.com/hurali97/status/2075536145509961971?ref%5Fsrc=twsrc%5Etfw) [Post on X](https://x.com/hurali97/status/2075536145509961971) We're also experimenting with a migration agent powered by [Apex](https://www.callstack.com/blog/introducing-apex-a-fast-specialized-model-for-react-native). Give it an existing screen or full bundle, and it produces a React Native and Expo build deployed to EAS, entirely in the cloud. It's early, but promising. I hope Hur ships something in this area soon. ### On-device AI and other experiments Local agents got hardware. Artur got Gemini Nano tool calling on a Pixel 9 through Google ADK, while Lech turned a dual-screen handheld into [Vault, a terminal for self-hosted AI agents](https://x.com/thymikee/status/2079896979702776157). > > dude just built a hardware e-ink terminal to access his [@openclaw](https://x.com/openclaw?ref%5Fsrc=twsrc%5Etfw) 🤯 [pic.twitter.com/kJBCcYVLVU](https://t.co/kJBCcYVLVU) > > > > — Michał Pierzchała (@thymikee) [July 22, 2026](https://x.com/thymikee/status/2079896979702776157?ref%5Fsrc=twsrc%5Etfw) [Post on X](https://x.com/thymikee/status/2079896979702776157) Elsewhere, Sniffler's parsing work became roughly 50% faster, cutting about a third from a typical run. We also started shaping an agent skill for assessing React Native migrations. Small experiments, useful sharp edges. ### Core React Native contributions Kamil posted a [CSS calc RFC](https://x.com/_ikswodarap/status/2079957753729044627) that brings familiar web-style calculations to React Native through a C++ implementation, without extra `onLayout` render passes. > > Just have opened a RFC for calc() support in React Native! > > > > Check it out and join the discussion: > > > > — Kamil Paradowski (@\_ikswodarap) [July 22, 2026](https://x.com/%5Fikswodarap/status/2079957753729044627?ref%5Fsrc=twsrc%5Etfw) [Post on X](https://x.com/_ikswodarap/status/2079957753729044627) He also pushed ResizeObserver forward through three core changes: [a new UIManager hook](https://github.com/react/react-native/pull/57721), [separate layout event emitters](https://github.com/react/react-native/pull/57722), and [the ResizeObserver implementation](https://github.com/react/react-native/pull/57723). Artur tested the SceneDelegate work on a physical iOS 27 device and contributed fixes for duplicate enums and cross-platform documentation differences in the C++ API codegen. [Riccardo has taken over the PR, which is now in review](https://github.com/react/react-native/pull/57700). We hope it lands in React Native 0.87. Apple is expected to require SceneDelegate for newly submitted App Store apps around the iOS 27 release, so the timing matters. ## **Blog** - [3 Minutes With an Agent, 9 Seconds on Replay](https://www.callstack.com/blog/3-minutes-with-an-agent-9-seconds-on-replay) by Michał Pierzchała: Agent-device v0.20 turns known mobile journeys into scripts that agents can replay in seconds without further model calls. This article reports the experiments behind that workflow, based on 20 repeated runs, 38 Maestro flows, and 36 repair trials. It maps out when a task calls for settled observations, deterministic replay, live execution, repair, or a full re-recording as an app changes. It also covers how to validate repaired flows, prevent plausible but incorrect actions, and measure performance under realistic load. - [Building Vault: A Dual-Screen Terminal for Self-Hosted AI Agents](https://www.callstack.com/blog/building-vault-a-dual-screen-terminal-for-self-hosted-ai-agents) by **Lech Kalinowski**: Vault is a dual-screen handheld terminal for controlling AI agents and locally hosted LLMs from a purpose-built device. This article explains how the prototype combines an ESP32-S3, OLED, and e-paper display with an OpenClaw agent and gpt-oss-120b served on a LAN host through TensorRT-LLM. It covers the interface architecture that separates live task status from completed output, the Python broker that keeps inference off the device, the hardware failures that shaped the build, and the security limits of remote command execution. - [A Step-By-Step Guide to Super App Development With Re.Pack 5](https://www.callstack.com/blog/step-by-step-guide-to-super-app-development) by **Nazar Sydiaha**: This hands-on guide shows how to turn two React Native applications into a minimal super app. Configure Re.Pack 5 and Module Federation, expose a mini app, register it in a host, and load it with React.lazy while handling loading and error states. - [An AI Agent for Measured React Native Performance Fixes](https://www.callstack.com/blog/an-ai-agent-for-measured-react-native-performance-fixes) by **Xavier Costa**: React Native performance work can produce misleading results when teams rely on a single profiler run. This article introduces Metrognome, an open-source AI plugin that turns proposed fixes into repeated, gated experiments on a live app. It explains how the tool combines device control, render analysis, runtime signals, and performance guidance, shares examples from Uphold, and shows how its Performance Map and CI workflows prioritize issues and create evidence-backed PRs. - [How we made Hybrid iOS apps compile 39% faster with React Native Brownfield](https://www.callstack.com/blog/how-we-made-hybrid-ios-apps-compile-39-faster-with-react-native-brownfield) by **Artur Morys-Magiera**: React Native Brownfield 3.10 brings iOS prebuilts into the packaging flow, removing a major compile step from hybrid app builds. This article explains how the integration works, what changed under the hood, and what our benchmark results show for React Native and Expo projects. - [Introducing Sniffler: Run only the E2E tests your changes can affect](https://www.callstack.com/blog/introducing-sniffler-run-only-the-e2e-tests-your-changes-can-affect) by **Szymon Chmal**: Sniffler helps JavaScript and TypeScript teams run only the E2E tests affected by a change. It traces import and export relationships and reduces unnecessary CI work without asking teams to give up test coverage. - [Mobile Module Federation With Re.Pack: When Runtime Delivery Is Worth the Complexity](https://www.callstack.com/blog/mobile-module-federation-with-re-pack-when-runtime-delivery-is-worth-the-complexity) by **Szymon Chmal**: This article explains what Module Federation means in React Native and how Re.Pack brings runtime-loaded JavaScript modules to mobile apps. It shows where this architecture can help, where packages or code splitting are the better choice, and what teams need to validate before adopting it in production. ## Podcasts and interviews - [React Native Ease: Platform-Native Animations Without the Extra Work](https://www.callstack.com/podcasts/react-native-ease-platform-native-animations-without-the-extra-work) - Łukasz talks with Janic from App & Flow about React Native Ease, Core Animation, Reanimated trade-offs, and performant UI animation in React Native. - [What AI Changes for Paywalls and Monetization](https://www.callstack.com/podcasts/what-ai-changes-for-paywalls-and-monetization) - Perttu Lähteenlahti (RevenueCat) talks about AI-generated paywalls, MCP tooling, monetization experiments, and what RevenueCat is building next across mobile and web. - [Where Reanimated Goes Next](https://www.callstack.com/podcasts/where-reanimated-goes-next) - A conversation with Bartłomiej Błoniarz on CSS animations in Reanimated, native primitives on iOS, shared element transitions, and the work still left on stability. - [Why Nitro Is More Than a Faster Bridge](https://www.callstack.com/podcast) - Marc Rousavy (Margelo) talks Nitro modules, native state, zero-copy data, Vision Camera, and why speed is only part of the story. - [Redraw for Effects Beyond Skia](https://www.callstack.com/podcast) - William Candillon talks about where Redraw fits next to Skia, from physics-based rendering to camera effects, AR, and new UI patterns. ## Webinars and live streams - [**Behavior Review for AI-Native Delivery**](https://www.callstack.com/events/behavior-review-for-ai-native-delivery) **by Michał Pierzchała and Piotr Miłkowski:** AI can increase code output faster than teams can verify the real app. See why behavior review matters before merge. - [**How We Built a React Native News Agent With Eve:**](https://www.callstack.com/events/build-a-react-native-news-agent-with-eve) See how we built and deployed an Eve agent that tracks React and React Native repositories, filters commit noise, and publishes daily Slack updates.\` - [How to Teach Your Own AI Model, Part 2: What Should You Build First?](https://www.callstack.com/events/how-to-teach-your-own-ai-model-part-2-what-should-you-build-first) Explore practical AI model training, from choosing a first project and preparing data to fine-tuning locally and evaluating synthetic data. - [**Building AI Agents With Eve: Tools, Skills, and Evals**](https://www.callstack.com/events/can-eve-make-ai-agents-easier-to-build-and-run) Learn how Eve helps structure AI agents with tools, skills, sandboxes, local models, HTTP channels, and evals. ## Events ### **Agent Conf** The speaker lineup for Agent Conf is getting stronger every week. So far, the program includes speakers from Amazon, Meta, Google, OpenAI, Curson, eBay, Postman, Allegro, AT&T, Expo, Epicweb.dev, Cognition, and more, across four tracks: Vibe Coding, AI Engineering, Orchestration, and AI Adoption. Alongside names we already shared, like [**Kent C. Dodds**](https://www.linkedin.com/in/kentcdodds/), [**Tejas Kumar**](https://www.linkedin.com/in/tejasq/), [**Nader Dabit**](https://www.linkedin.com/in/naderdabit/), [**Kitze**](https://www.linkedin.com/in/kitze-r-9a286b303/), [**Levan Kvirkvelia**](https://www.linkedin.com/in/levan-kvirkvelia-2914a8120/), [**David Mokos**](https://www.linkedin.com/in/david-mokos/), [**Nishant Gupta**](https://www.linkedin.com/in/nishantgupta-ai/), [**Sven Großmann**](https://www.linkedin.com/in/sven-gro%C3%9Fmann-26387311b/), [**Sujith N S**](https://www.linkedin.com/in/sujith-n-s-600641a1/), and [**Ruben Casas**](https://www.linkedin.com/in/ruben-casas-17100383/), the program now also includes [**Carl Ross**](https://www.linkedin.com/in/carljross/) from OpenAI, [**Nick Miller**](https://x.com/nickwm) from Cursor, [**Anisha Malde**](https://www.linkedin.com/in/anishamalde/) and [**Kourtney Meiss**](https://www.linkedin.com/in/kourtney-m-a59792a6/) from Amazon, [**Piotr Karwatka**](https://www.linkedin.com/in/piotrkarwatka/) from Open Mercato, [**Ivana Huckova**](https://www.linkedin.com/in/ivana-huckova/) from Grafana Labs, [**Jakub Kuzimski**](https://www.linkedin.com/in/jakubkuzimski/) from Allegro, [**Zuzanna Skoczylas**](https://www.linkedin.com/in/zuzanna-skoczylas-b22339151/) from Sierra, [**Leonel Candido da Silva**](https://www.linkedin.com/in/leonelcandidosilva/) from eBay, [**George Nikolaropoulos**](https://www.linkedin.com/in/gnikolaropoulos/) from Natech, [**Anjaneya Sastry Kappagantu**](https://www.linkedin.com/in/anjsastry/) from AT&T, [**David Budáč**](https://www.linkedin.com/in/david-budac-98889392/) from ČSOB, and [**Susmita Singh**](https://www.linkedin.com/in/susmita-singh-564624101/) from Google. A few of the newly announced talks that stand out are Agents FC: Harness Your Agents Into a Winning Team, When "Vibes" Aren't Enough: Observability for AI Agents, The AI Agent Iceberg, and Beyond the Prototype: Building Resilient, Non-Hallucinating AI Agents at Scale. The [**agenda**](https://www.agent.sh/agenda) is now live, so you can see when each session is happening and start planning your two days at Agent Conf. Curious about the talks? Learn more here: [https://www.agent.sh/#speakers](https://www.agent.sh/#speakers)‍ Join us in Warsaw: [https://www.agent.sh/tickets](https://www.agent.sh/tickets) ### **Agentic AI in the Wild - Boston Dev Community Meetup** We’re heading to Boston on August 18 for Agentic AI in the Wild - Boston Dev Community Meetup. The evening is focused on what happens after the demo stage, when AI agents start touching code, CI/CD pipelines, developer platforms, and production workflows. The lineup includes [**Matt Hargett**](https://www.linkedin.com/in/matthargett/) from Callstack, [**Nnenna Ndukwe**](https://www.linkedin.com/in/nnenna-ndukwe/) from Qodo, and [**Rafael Mendiola**](https://www.linkedin.com/in/rafmendiola/) from [okthink.ai](http://okthink.ai). You’ll learn _How to Embed Code Quality into SDLC_ from Nnenna Ndukwe, hear about _Large Codebase Migration with Loop Engineering_ from Rafael Mendiola and listen to Matt Hargett's talk,_Fuzz First, Farm Later: Local Validation Before Costly Cloud_. If you’re in town, come join us for the discussion: [Agentic AI in the Wild - Boston Dev Community Meetup.](https://www.callstack.com/events/agentic-ai-in-the-wild---boston-dev-community-meetup)