What happened
In an engineering post published November 10, 2026, Shopify describes a rewrite of its mobile app by roughly 30 engineers. The app has more than 300 screens, home- and lock-screen widgets, an Apple Watch app, complications, and Siri Shortcuts. The rewrite moved the app from React Native to native Swift and Kotlin, and Shopify says agents wrote most of the code. The post frames the project as the integration of several efforts Shopify had described separately: its move to native mobile, Helix, a remote development environment for agents, and Reflex, an architecture for mobile apps built with agents.
Shopify says a small group first prepared the architecture, tooling, and feedback loop before migrating screens. The resulting system made business logic runnable without a device, offered a command-line interface to load screens quickly, and used a shared architecture for iOS and Android. Examples, Markdown guidance, naming and state conventions, lint rules, test fixtures, and CI failures made the intended patterns visible to agents. Shopify contrasts that with the previous loop of building the app, opening an emulator or simulator, navigating to a screen, inspecting screenshots, and iterating—a process it says could take several minutes per iteration. In the new setup, an agent could load a ViewModel from the terminal, trigger actions, inspect state, and run deterministic checks in milliseconds.
The engineers worked in pairs, with each pair owning an area across both platforms. Shopify says each pair steered multiple agents, and agents continued working overnight and on weekends, leaving queues of pull requests for morning review. The company deliberately favored consistency between platforms over platform-specific architecture. The same screen exposes the same state, actions, and behavior on iOS and Android; a pull request can contain both platform changes; and CI fails when the platforms diverge in ways the architecture disallows. Shopify says this made examples reusable and allowed engineers familiar with one platform to review work on the other.
Shopify says it tried spec-driven development but found that screen specifications quickly became stale. For this migration, the existing React Native app served as the behavioral reference. Helix reads that reference, breaks migration into small checkpoints, and checks each checkpoint against the original app. The company says the old screen captured states, actions, edge cases, loading and empty states, and navigation. It draws a distinction between migration and new product work: when no existing screen provides a reference, the approach may need to change. Across the workflow, humans retained ownership of architecture, product behavior, and final judgment about what was fit to ship.
What the archive adds
New Runtime's earlier coverage has examined Shopify's task-expert loop in How Shopify turns open models into task experts (October 22, 2026), where narrow agent workloads—not general model rankings—were the optimization target. This is a new case with a related operational emphasis: Shopify now describes how architecture and verification enabled agents to execute a large mobile migration.
Trend signal
This supports verification-bandwidth-is-the-scarce-resource: Shopify says implementation stopped being the slowest part, while review, integration, decisions, and CI became more consequential. The account is one company's retrospective rather than a general measurement, but its workflow details make the bottleneck shift concrete.
Why now
Shopify presents the migration as a consequence of agents producing code faster than the surrounding development process could originally validate it. Its response was to redesign the workflow before scaling the rewrite, rather than treat faster code generation alone as the objective.
Why it matters
For teams adopting coding agents, Shopify's account shifts the design question from “How many lines can an agent produce?” to “How can a change demonstrate that it is correct and reviewable?” The useful investments in this case are structural: runnable business logic outside a device, a shared cross-platform shape, small checkpoints, and CI rules that enforce architecture rather than relying on prompts or reviewers to notice every deviation.
The existing product can also be a stronger migration specification than a document that must be kept synchronized. That does not make a live reference sufficient for new features, where behavior has not yet been implemented. Teams should distinguish work that converges on an existing product from work that must define new behavior, then choose their validation method accordingly. Shopify's experience suggests agent capacity has limited value unless machines, tests, review ownership, and integration throughput can absorb the output.
Limits and uncertainty
This is Shopify's account of its own project, not an independently audited productivity study. The post gives a team size, timeline, and app scope, but does not establish a comparable baseline for how long the rewrite would have taken without agents or quantify quality, defect rates, or the total review effort. Shopify's description also reflects its internal tools and conventions; the results may depend on access to the legacy app, a team able to invest in shared architecture, and sufficient CI and device capacity. The migration's reported completion does not by itself show how the approach performs on greenfield features, which Shopify identifies as a different challenge.
Skill to build now
For a migration, choose one existing screen and turn its current implementation into a repeatable acceptance loop: identify its states and actions, make the relevant logic runnable without a full device session, and split the replacement into checkpoints that can be compared with the reference. Add a CI check for one architectural invariant before expanding the pattern. Shopify's account makes this concrete for work where a working product already serves as the behavioral target.
Reader takeaway
The notable part of Shopify's 12-week rewrite is not simply that agents wrote most of the code. It is that the company made architecture observable, checks fast, and work consistent across platforms so humans could review what agents produced. If implementation gets cheaper, teams still need a system that decides what should ship—and catches when generated work drifts from it.