AI Agents Write the Code. How Do Developers Stay in Control? | React Universe On Air

Reliable AI workflows depend on the context, tools, and feedback you give your agents. Harness engineering brings those pieces together, but the question is, how do you choose them, and how does your role change when AI handles more of the coding? Amazon’s Anisha Malde joins Łukasz Chludziński for this Coffee Talk episode of React Universe On Air, recorded at Agent Conf 2026 in Warsaw. They discuss harness engineering through examples from Amazon Devices Builder Tools (ADBT), Amazon’s Vega OS, and React Native, covering Fire OS-to-Vega migration, reusable agent skills, and the technical judgment developers still need. Tune in and learn: 🔹 How prompt engineering, context engineering, and harness engineering fit together. 🔹 How Amazon combines Model Context Protocol (MCP) tools, agent skills, and workflows for device development. 🔹 How platform expertise becomes reusable guidance, and where platform-specific knowledge is still needed. 🔹 What developers risk losing when they work further from the implementation. 🔹 How task breakdown, specifications, and CI/CD translate to agent workflows. 🔹 Keeping control of the user experience as AI takes on more implementation work. Subscribe to React Universe On Air for conversations about React, React Native, and how modern teams build and ship software.

AI can generate code faster than your team can review, test, or release it. That extra output can turn the next stage of software delivery into a bottleneck. Matt Hargett, Consulting Principal Engineer at Callstack, joins Łukasz Chludziński to examine how AI agents change the software development lifecycle (SDLC). Drawing on nearly 30 years of experience, from QA at McAfee to agent workflows today, Matt explains how to find the current constraint and anticipate where work will pile up next. In this episode, we discuss: ➡️ Why an on-time release can still lose money through support and infrastructure costs ➡️ How to give agents goals, context, and explicit quality expectations ➡️ How coverage-guided fuzzing can become an overnight testing and repair loop ➡️ How deployment bottlenecks grow release batches, and where phased rollouts and over-the-air updates fit ➡️ When to rethink Jira tickets and handoffs between teams ➡️ How to measure whether AI adoption improves delivery through support load, uptime, and production feedback Which part of your delivery process would struggle if coding got twice as fast? Share your experience in the comments, and subscribe for more React Universe On Air conversations. Check out episode resources on our website ➡️ https://clstk.com/4yF4yjl Catch more React Universe On Air episodes 🎧 https://clstk.com/4B1G9q4 Sign up for our newsletter ✉️ https://clstk.com/3VsrLXB Follow us on X 🐦 https://x.com/callstackio Chapters: 00:00 Faster coding, same deployment bottleneck 00:54 Meet Matt Hargett 02:07 QA at McAfee in 1997 07:19 Product economics and estimation 11:49 Mapping the SDLC in the AI era 16:25 Shipping, quality, and operating costs 22:34 Delegating to agents: goals and guardrails 25:21 Coverage-guided fuzzing explained 28:06 Overnight testing and repair loops 30:06 How agents change engineering roles 36:55 Do agents still need Jira tickets? 44:16 Using customer pain to prioritize improvements 49:41 Finding your next delivery bottleneck 54:07 Phased deployments, staging, and telemetry 57:06 Examining how teams really work 01:01:18 Improving the SDLC incrementally 01:03:06 Measuring outcomes and getting outside help 01:06:52 Closing thoughts

Many AI features in web apps send the user's audio, images, or text to a server and wait for the result. That adds network latency, moves data off the device, and ties the experience to a connection. Transformers.js gives developers another option: run the model in the browser. In this episode, Mike Grabowski speaks with Nico Martin, Open Source ML Engineer at Hugging Face, about what browser AI can do today. They break down the Transformers.js pipeline API, ONNX Runtime, WebGPU acceleration, model downloads, browser caching, CPU fallbacks, and performance across devices. Nico explains when local inference can beat a server round trip, why model size and hardware variation shape the user experience, and how browser code can handle parts of an agent workflow. He also shares early work on structured output and a custom WebGPU inference engine, where current experiments point to 5x to 10x speedups. For developers deciding which AI tasks belong in the browser and which need the cloud, this episode maps the tradeoffs through speech recognition, background removal, embeddings, local tool calling, and on-device agents. Check out episode resources on our website ➡️ https://clstk.com/4zMwzq9 Catch more React Universe On Air episodes 🎧 https://clstk.com/4zNCV8W Sign up for our newsletter ✉️ https://clstk.com/4h2b8Ks Follow us on X 🐦 https://x.com/callstackio Chapters: 0:00 Welcome to React Universe On Air 1:04 Nico Martin's work at Hugging Face 2:04 What Transformers.js is 3:30 Browser AI beyond large language models 6:00 Bringing Python ideas into a JavaScript API 10:08 How the Transformers.js pipeline API works 14:23 Choosing compatible models and architectures 15:46 Why Transformers.js uses ONNX Runtime 19:01 CPU inference, WebGPU, and browser coverage 22:57 Model downloads, browser caching, and UX 28:28 Startup time and performance across devices 30:57 Why Transformers.js focuses on the browser 34:07 What developers misunderstand about browser AI 38:44 Building AI agents in the browser 43:19 Local fallbacks and hybrid AI 44:28 The browser AI roadmap 47:04 Hugging Face's custom WebGPU inference engine 50:57 Why speedups matter on constrained hardware 52:07 Nico's favorite browser AI use cases

VisionCamera v5 was five years in the making, and in this episode, Marc Rousavy joins Mike Grabowski to explain what changed, why v5 is more than a version bump, and how Nitro reshaped the foundation of one of the most widely used camera libraries in React Native. They cover the bridge and JSI limitations that pushed VisionCamera toward a new architecture, the move to Nitro Modules, and the API decisions behind v5. The conversation also gets into the new constraints model, output objects, in-memory photo handling, barcode scanning, migration from v4, and the testing work needed to make camera features hold up across real devices. If you build React Native camera apps, this episode is a detailed look at the engineering behind VisionCamera v5 and the decisions behind making complex camera features easier to ship and maintain. Check out episode resources on our website: https://clstk.com/3Rj0dlN Catch more React Universe On Air episodes: https://clstk.com/45ypLys Sign up for our newsletter: https://clstk.com/4bktu6f Follow us on X: https://x.com/callstackio Chapters: 0:00 Welcome to React Universe On Air 2:07 What VisionCamera is today 4:51 Barcode scanning and the new Data Scanner library 7:25 How VisionCamera helped shape Nitro 10:55 Frame processors, C++ glue, and the old JSI era 18:56 What is actually new in VisionCamera v5 21:45 The constraints API explained 30:12 Why third-party camera apps still struggle 33:29 The new output system and custom outputs 37:31 Imperative API, hooks, and components 39:01 Migrating from v4 to v5 43:25 AI, agents, and building VisionCamera v5 45:33 Testing, stability, and what comes next 48:25 Advice for teams building camera apps 50:50 Why web support is not on the roadmap

Some animations do not need a full animation engine. Sometimes you just want a view to fade in, translate, or move slowly without burning frame budget every time the UI thread gets busy. Łukasz Chludziński talks with Janic Duplessis from App & Flow about React Native Ease, a small animation library built around platform-native primitives. They cover early React Native days, why Reanimated is still the right tool for complex interactions, and where a simpler React-like API can make sense. Key takeaways: - Why React Native Ease was created and what problem it solves - How Core Animation can run animations outside the UI thread on iOS - Why simple transitions do not always need Reanimated’s full flexibility - Where Reanimated remains the right tool, especially for gestures and complex interactions - How animation performance relates to frame budget, not just visible frame drops - Why React Native’s maturity shifts more innovation into third-party libraries - How focused open-source tools can push the whole ecosystem forward Guest introduction: Janic Duplessis is a developer at App & Flow and the creator of React Native Ease. He has been working with React Native since the early iOS-only days and has contributed to React Native Core, including work around animation internals and native-driver concepts. Check out episode resources on our website 📚 Catch more React Universe On Air episodes 🎧 https://www.callstack.com/podcast Sign up for our newsletter ✉️ https://www.callstack.com/newsletter Follow us on X 🐦 https://x.com/callstackio

Rudrank Riyam joins React Universe On Air to talk about App Store Connect CLI, an open-source tool for automating App Store Connect workflows without living in the dashboard. What started as a way to pull TestFlight feedback turned into a broader interface for metadata, submissions, localization, analytics, reviews, and release operations. Mike Grabowski and Rudrank discuss why these workflows are a strong fit for CLI automation, why Go became the right language for an agent-built tool, how teams are using it in CI/CD and agent workflows, and what changes when multiple coding agents build against the same codebase. They also get into code structure, command design, agents.md, public vs. private App Store Connect endpoints, and what changed after Rork acquired the project while keeping it open source. What we cover: - Why App Store Connect workflows are a good fit for CLI automation - How the tool handles TestFlight feedback, metadata, localization, analytics, reviews, and submissions - Why Rudrank chose Go over Swift for an agent-built CLI - What he learned from running multiple agents against one open-source project - How clear structure and command naming make a codebase easier for both humans and agents - What Rork’s acquisition means for the project going forward Guest: Rudrank Riyam is a software engineer at Rork and the creator of App Store Connect CLI. He works at the intersection of Apple development, release tooling, and agentic workflows.

Zack Ebenfeld joins Michał Pierzchała to talk about AI-assisted coding, testing, QA, code review, and why verification matters more than ever in mobile workflows. It is a practical conversation about trust, feedback loops, and how engineering work is shifting from pure implementation toward judgment, architecture, and product thinking.

What happens when you combine voice agents and teddy bears? A hands-on look at local-first systems, voice AI, and security lessons from poking real apps. Kraen Hansen and Ola Desmurs-Linczewska talk about Software, AI Agents, parenting, stoicism and more!

Vercel’s v0 iOS app is built with React Native, but it looks and feels like a fully native iOS product. In this episode of React Universe On Air, we go deep into how the v0 mobile app was built: from choosing Expo and going iOS-first, to writing native modules, fixing low-level React Native issues, and upstreaming improvements back to the ecosystem. This is a hands-on engineering conversation with Fernando Rojo, Szymon Rybczak, and Oskar kwaśniewski about real trade-offs, real constraints, and what it takes to ship a polished app under real deadlines. Chapters 00:00 Welcome to the React Universe 00:50 Meet our guests 02:42 v0 app overview 03:45 Iterating to greatness 05:07 Choosing Expo and initial tech decisions 07:55 Code sharing with v0 website 11:03 Open Source contributions 13:20 React Native Core fixes 15:40 Android version? 18:42 Modal issues we fixed 23:20 Working with React Native Core 26:16 Balancing deadlines and root cause analysis 27:23 Native Modules and custom solutions 29:16 Minimalistic UI challenges 30:44 Floating Composer concept 33:41 Feature flags and testing 36:48 WebView swiping and custom forks 45:16 Launch experience and security concerns 50:17 Final thoughts

This year, React Universe On Air is doing something different for Christmas. No guests, demos, or framework updates. Instead, Ola Desmurs Linczewska tells a Christmas story. A (Secure) Christmas Carol is a holiday fairytale for JavaScript and React Native developers. Set on Christmas Eve, it follows Npmezer Scrooge, a senior engineer who believes deadlines matter more than holidays, and warnings can always wait. Over the course of one long winter night, he’s visited by unexpected guides who show him what his choices look like across time: the optimism of the past, the quiet damage of the present, and a future no one wants. All technical details, real-world security issues, and references mentioned in the story are explained in the show notes 📚  https://clstk.com/4pEVYwv We’ll be back in January with regular episodes. Until then, stay kind, stay curious, and have yourself a very merry Christmas.