On September 10, Shopify’s engineering blog published “Native is now the future of mobile at Shopify,” announcing that every mobile app the company ships will be rebuilt in Swift and Kotlin — reversing the 2020 decision that made React Native Shopify’s mobile standard. The Shop app is already done: six engineers, twelve weeks, from proof of concept to a shipped native build in the App Store and Play Store. iOS cold start fell from 3,200 milliseconds to 2,466. Android fell from 4,433 to 2,233.

A Hacker News thread on it is sitting at over a thousand points and three-quarters of a thousand comments, most of them arguing about frameworks. They are arguing about the wrong thing.

Shopify’s own explanation is more interesting than the fight it started. The company says coding agents changed the economics: LLMs are good enough at writing Swift and Kotlin with the React Native version as a reference that building the app twice is now cheaper than maintaining one app that runs adequately on neither platform. That is a real argument and it deserves a real hearing. It is also being read far more broadly than the evidence supports.

The Agents Ported a Finished App. They Didn’t Design One.

Read the migration writeup closely and the task Shopify handed its agents comes into focus. They were not given a product spec and a blank page. They were given an app that had already been shipped, measured, debugged, and lived in for years, and asked to reproduce it in another language.

That is a port. A port with a ground-truth reference implementation is the single easiest job in software. The hard parts — discovering that your data model was wrong, that your navigation graph fights your deep links, that your background sync was a lie — have already been solved once, in public, by humans. Transliterating a working system into a second language is precisely the pattern-matching these models are good at.

Which is a fine reason to do it. It is not evidence that agents can build mobile apps. Ask again when somebody hands one a blank page and a vague product brief and gets a shipped app out the other end.

4,433 Milliseconds Is Not a Framework Benchmark

The Android number is the one circulating, and it proves the least. Shopify’s post says the rewrite gave the team “a clean slate to rebuild in the best way possible without any of the previous constraints.” That is a candid sentence, and it quietly detonates the comparison.

What was measured is not Swift against React Native. It is a new codebase against a five-year-old one. Any team that throws an app away and starts over gets a speedup, in any language, with or without an LLM, because five years of accreted workarounds, dead feature flags, and dependency debt do not survive a rewrite. The 50% is mostly the cost of the old app finally being paid off. Shopify could have rebuilt in React Native from scratch and booked a meaningful share of it.

Note the timeline, too. In January 2025 the same team published “Five years of React Native at Shopify,” reporting that the framework’s future was bright and that the company planned to keep investing. Twenty months later the future is Swift. Framework decisions are made by people with budgets and quarterly goals, not by benchmarks, and the sooner everyone stops treating a company blog post as an empirical result, the better these arguments get.

The Abstraction Layer’s Real Job Was Headcount

React Native was never adopted because it was fast. It was adopted because it let one team cover two platforms. The parity problem is a staffing problem wearing a technical costume, and the abstraction layer was the costume.

Shopify just showed that six people and a set of agents can absorb parity work that used to justify a mobile platform organization. Hold that next to Tobi Lütke’s memo telling staff that teams must demonstrate why the work cannot be done with AI before asking for more headcount, as Fortune reported. Read it as productivity policy and it is anodyne. Read it beside this migration and it is a hiring freeze with better manners.

A platform engineer at a rival commerce company put it more bluntly in a Slack message sent while a build was running: “Everyone’s arguing about Swift. Nobody’s asking who was maintaining the bridge.”

That is the part that should bother people who have been telling themselves the AI headcount squeeze is a story about badly run companies cutting their way to a number. Shopify is not badly run. It is one of the better-run software companies alive, and this is what the squeeze looks like when it is done well: a small team, a coherent technical rationale, and a job category quietly deleted between two blog posts.

The boosters should be just as uneasy. The flagship example of agents doing serious engineering work this week is a transliteration job with an answer key. That is the easiest item on the menu, and it is being served as proof of the whole restaurant.

Sources