Hey,
It's Michał with the October edition of the Callstack newsletter, catching up on September and the first half of October.
Apex has launched, we shipped our first collaboration with Margelo, and agent-device is powering e2e and T3 Code - wild! Outside the Incubator, React Native Web has a new maintenance plan, Shopify is changing its mobile stack, and there's a Next.js security update to install.
It's a packed one. Let's go!
Community highlights
- Next.js shipped an urgent security update: The October 14 release fixes two critical remote code execution vulnerabilities and a high-severity denial of service issue. Check the advisory and upgrade to the patched version for your release line: 16.4.1 or 15.5.28.
- React Native Web has a renewed maintenance plan: Mathieu Acthernoene, Nicolas Gallagher, Expo, and community contributors are picking up active development, with a move to the React Foundation planned. TypeScript, ESM, and closer alignment with React Native are on the roadmap. The modernization work lives on a separate branch for now, feel free to contribute your patches.
- Airbnb-moment is backShopify is moving back to Swift and Kotlin: The team says coding agents changed the cost of maintaining two implementations. Its Helix write-up explains the small checkpoints, tests, visual checks, and human reviews behind the migration. There's also a practical consequence for React Native users: read the maintenance plans for Skia, FlashList, and Restyle if your app depends on them.
- Reanimated 4.7 enables its new layout animation engine by default: It also adds animated
backgroundImagestyles and more Android CSS transition support. Shared element transitions remain experimental. The 4.7.1 follow-up includes stability and crash fixes, so check the patch release when planning an upgrade. - React DevTools 8 makes Suspense easier to inspect: The Suspense tab is enabled by default, and selecting a DOM node can reveal the matching React component. The old Timeline profiler tab has been removed; use the browser's Performance panel and React tracks for that workflow.
- TanStack Start 1.0 is stable: The full-stack framework combines typed routing, server rendering, streaming, and server functions with a client-first application model. After a long prerelease period, there's now a stable API commitment for teams considering it for production.
- TesterArmy launched e2e for web and mobile apps: Write predictable parts of a test in code, then use natural-language actions and assertions where you need an agent. Playwright drives the browser and agent-device drives iOS simulators and Android emulators. Supports EAS Simulators and Limrun for remote devices. It's open source and still heading toward 1.0, so expect APIs to evolve. Expo's explainer shows how it fits into a verification loop.
Callstack Incubator highlights
Apex
I’m happy to finally say it: Apex is generally available.
It's our coding model for everyday agent work, with particular strength in React Native, React, and Next.js. You can use the direct API or Vercel AI Gateway, and connect it to tools such as Claude Code, Codex, Cursor, and OpenCode with npx @callstack/apex@latest init. The model currently builds on Qwen, with our own domain data, training, and evaluation pipeline.
If you tried Pixel Canary on Vercel AI Gateway, you were already using Apex (but it was painfully slow in the first few days; sorry, we overestimated our infra capacity). The stealth preview served more than 200 billion tokens over five days and gave us plenty of infrastructure work before the public launch. We’re now able to easily serve Apex at 100-200 tokens per second under 1s latency.
On Vercel's 31-task Next.js evaluation, Pixel Canary completed 28 tasks without the docs and 30 with the docs in context. Apex also scored #2 in Fixing React category and fifth on ReactBench's best-per-model leaderboard with its medium-effort run. Not too bad for a bunch of nerds at a European company, that’s not even technically a lab. Different benchmarks test different work; try it on tasks from your own repository.
Direct API pricing is $0.50 per million input tokens, $0.20 for cached input, and $3.00 for output.
We host Apex for all Callstack and Margelo employees giving them unlimited access. We can also host it (or any other open-weight model really) for you, with decent per-watt performance. If your tokens get too expensive, send me an email.
Mobile Dev, our first collaboration with Margelo
We’ve recently shared that Margelo is joining Callstack. Now, we have our first joint release to show you. Laurids and Hanno built Mobile Dev for Codex, bringing live iOS and Android previews, screen annotations, native and JavaScript logs, and performance profiling into the workspace beside your chat.
You can point at a screen region and tell the agent what to change, inspect a crash, or compare performance recordings before and after a change. Built-in agent-device support lets the agent operate the app alongside you. It works with React Native, Expo, SwiftUI, and native mobile projects.
The initial preview is available through the installation instructions on GitHub. Watch Laurids's demo, which is probably the quickest way to see why we built it.
agent-device
The e2e launch was also a collaboration for us. We worked with Oskar (who was on our team a while back!) on the mobile integration, so tests can combine agent-driven interaction with deterministic checks through agent-device.
We also added iPhone Duo simulator controls, letting agents change the folded state through the tool. Android snapshots gained more information about controls, including placeholder text and checked state, alongside continued work on text entry and interaction reliability. And as always, I’m doing tonnes of refactoring to allow any cheap agent be productive in the codebase, recently dealing with CRAP (Change Risk Anti-Patterns). You should really learn about this metric.
Mike paired agent-device with Jev to explore a different kind of QA loop: Jev chooses from structured actions, and agent-device reads and operates the app. The published proof-of-concept run took 14 seconds and cost $0.0023 in model inference. e2e made the API easier so you can mix decision model with LLMs.
Appduct
Speaking of Jev, we wanted to make it play Flappy Bird at Agent Conf. But guess what, agent-device is too slow for that (even when using some private API shenanigans).
Coincidentally, Szymon recently shipped Appduct, which is a reincarnation of Cordierite, which exposes functions in your running app to agents and tests through MCP. This allows Jev and anyone else call APIs exposed through the app, such as floppy_tap in under 16 ms (yes, milliseconds!) and control the game real-time.
We’re in post-Shopify migration world, so Appduct, of course, supports native Swift, Kotlin, and also our beloved React Native. Szymon's demo shows the idea in practice: let the app expose the operations your tests keep repeating.
Rozenite
Rozenite 2.5 continues the work on developer tools for React Native and Lynx.
Initialization now detects Lynx projects. Network capture has been rebuilt around an XHR-first recorder, and the Redux plugin can filter actions with allowlists and denylists. Agent commands gained implicit sessions and consistent, bounded list responses, helping agents inspect state without pulling an entire history into their context.
The release also includes a redesigned website and refreshed DevTools panels.
React Navigation
Satya reworked navigation history handling for tabs and drawers, and added tabBarRepeatedPressBehavior to control scrolling and popping when someone presses the active tab again.
The latest React Navigation releases also improve lazy screen loading, native tab customization, header accessibility, and animation behavior under Strict Mode. Small changes, but these are the interactions people repeat throughout the day. We also published an AI usage policy. Contributors remain responsible for understanding their changes and taking part in the review.
Elsewhere in our navigation tooling, the standalone react-native-bottom-tabs library is getting ready for foldables with Davyd’s contributions
and Thiago created a new library react-native-arrangement-view.
React Native community is ready for iPhone Duo before it’s physically available.
React Native core contributions
Marcin Szalski has picked up the CSS work from Kamil and is pushing clip-path forward through separate PRs for JavaScript support, iOS, and Android. All three are still open.
He is also implementing and testing backdrop-filter on both platforms, including an iOS demo running in RNTester. These features are still in development; we'll share more as the work gets reviewed and lands.
A peek at Apex Sentinel
We're starting to test Apex Sentinel, our code review agent, on public repositories such as React Native Paper.
The team is evaluating its reviews against real pull requests, including React Native Paper changes. It's early; we'll share more when we have results to discuss.
Worth reading and watching
Blog
- How to Maximize AI SDLC? Measure the Outer Business Loop, Not the Inner Developer Loop by Matt Hargett: If you're measuring AI-assisted development by how quickly an agent finishes a task, Matt has a broader loop for you. The Outer Business Loop follows a change from the initial hypothesis to its effect on users and the business. His online platform optimization example covers checking who actually saw the feature, making sense of mixed results, and counting the full cost of an improvement.
- Building Reliable AI Figma-to-Code Loops by Vitaly Kuprin: Your agent can match the Figma screenshot and still ignore your component library. Vitaly shares how his team got agents to reuse existing components, blocked unsupported imports, and compared the rendered screens with the design. Some screenshot checks turned out to be misleading, too. He covers what caught those and where visual verification falls short.
Interviews and live streamS
- Inside Cursor, Colossus, and Grok Bot: Nick Miller joins us to talk about Cursor's move to SpaceXAI, what Colossus changed about model training, and how the browser experience behind Grok Bot works.
- Who Tests AI-Written Code?: Szymon Rybczak walks through how TesterArmy uses agents to test web and mobile apps, puts together test plans for pull requests, and fits into the rest of your QA workflow.
- Can Engineers Trust AI Agent Fleets?: Nishant Gupta from Meta talks about running fleets of AI agents and reusing the control planes behind them. You still need to check what the agents actually did, especially as you hand them more work.
- What Happens When AI Agents Have No Token Budget? OpenAI's Carl Ross talks about scaling coding agents when you stop giving them a token budget. That includes competing code changes, continuous cleanup, and the limits you still run into.
- State Machines for AI Agents: David Khourshid walks through using state machines to keep agent workflows under control, including human approval steps, tracing, and what to do when a model returns something invalid.
- AI Coding Agents and Hermes: Tzvetan Mikov talks about coding agents, context switching, and getting down to low-level engine optimization. Specs and tests still need a review, too. He also shares his plans for Hermes X.
- Can Jev Make AI Agent Decisions Faster? A closer look at Jev's structured, probability-based decisions: how they work alongside LLMs, whether they can make agent decisions faster, and whether you can get a similar approach running locally.
Upcoming events
See you in Amsterdam
On October 22, we're teaming up with JetBrains and Expo for From Design to Shipped: A Modern App Development Meetup at the JetBrains Amsterdam office. We'll cover how to use AI agents across web and mobile development, from design and code to checking that the app works.
I'll show a workflow for turning mobile designs into working features with a lower-cost model. Cedric van Putten will demonstrate Expo's verification loops, and Craig Walker and Kasper Peulen will discuss how WebStorm and Storybook are adapting to agents.
Join us for the talks, pizza, drinks, and a conversation about what it takes to ship the result. It's free to attend; please register in advance.
AI Coding Summit Berlin: save 15%
Further ahead, AI Coding Summit Berlin takes place in Berlin and online on December 4 & 7, 2026. The program has 40+ talks and workshops on agentic programming, spec-driven development, AI code review, evals, and agent security, with speakers from Microsoft, Hugging Face, AWS, and more.
Use code AGENT for 15% off.
That's it for October! Apex is out, our mobile tooling keeps growing, and there's plenty to try before the next edition.
If you're around, come say hi in Amsterdam. You can also join AI Coding Summit Berlin in person or online. I'd love to hear what you're building with all this.
See you out there,
Michał
