Shopify's Helix feeds its native rewrite to agents one screen at a time

Handing a model the whole React Native codebase and asking for Swift is the version everyone tries first, and Shopify tried it. Director of engineering Mustafa Ali describes the output as slop, and more planning did not rescue it: even with the model writing specs and task files before touching code, it still produced too much code engineers would not want to maintain, The New Stack reports ^[17][18].

Helix is what they built instead. It works one screen at a time and breaks each screen into smaller chunks rather than recreating everything at once. It compares the result against the existing app as it goes, separate agents look for problems before a human signs off, and the review feedback is retained so earlier findings inform later work ^[8]. Shopify's CEO had publicly threatened to ban Claude Code over engineers shipping agent-generated code without enough review ^[25], and The New Stack reads Helix the same way, treating the agent's output as something that must be verified before anyone relies on it ^[9].

The load-bearing input here is the app that already works, not the model performing the rewrite. In Shopify's prototypes, agents implemented a feature on Android using the iOS version as the reference, and the reverse ^[6]. One engineer spent a week on agent-assisted prototyping and got most of the Shop app ported from React Native to SwiftUI, with a result that was not production ready ^[15]. Both runs had a working implementation to diff against. Agents also let engineers contribute on platforms they did not know well, because the agent handled more of the platform-specific work ^[7].

Six engineers over 12 weeks is about 72 engineer-weeks to take Shop from proof of concept to a store release ^[4][19]. The merchant app is a different size: 300-plus screens, home and lockscreen widgets, an Apple Watch app, Siri Shortcuts ^[11], and it relies heavily on deep iOS platform integration ^[13]. The other number in circulation is Jarred Sumner's port of the Bun runtime from Zig to Rust, which The New Stack reports as 64 parallel Claude instances, roughly a million lines in 11 days, an estimated $165,000 of API cost, with the existing test suite already written and passing on all six supported platforms before the merge ^[21]. That is near 17 cents per line, or about $15,000 a day ^[24].

"LLMs changed one of the core assumptions behind our 2020 decision," wrote Ali ^[16]. Shopify's own post goes further, saying agents were "no longer just helping us write code faster. They were capable of making us question whether building software twice still meant doing twice the work" ^[14]. The company is also explicit about what has not changed: native still means building and maintaining two platforms, and that cost has not disappeared, it has stopped being the deciding factor it was in 2020 ^[10]. What comes off the other side is spend Shopify was already making, on React Native performance work, framework fundamentals, and keeping up with framework updates and external dependencies ^[12].

In May, Simon Willison wrote about meeting an engineer whose company had used coding agents to combine its old iPhone and Android apps into a single React Native app; Willison asked why they would consolidate when agents were making separate codebases easier to maintain, and the engineer said React Native did what the company needed and they could always move back to native ^[22]. The same agent tooling that makes native rebuilds cheaper also makes maintaining a shared codebase cheaper, so if your shared-codebase justification was priced when the second build meant a second team, the number is stale, and re-running it may still land on one codebase. The New Stack also reports that agents can create more work elsewhere in an engineering organization even as they make some development tasks much faster ^[23].