Shopify's Helix has LLMs port a 300-plus-screen app one reviewed checkpoint at a time

Shopify summed up the design in one line. "Helix builds a loop where an imperfect attempt cannot move forward until it becomes a good result," the company wrote ^[18]. When a gate fails, the agent takes the feedback, fixes the code and runs the check again ^[14]. It may retry as many times as it needs to. It may not override a failed check because it thinks the result is good enough ^[14].

Most of the human judgement goes in before any code exists. The first checkpoint is usually the screen skeleton and the second is a deliberately small section. Later ones grow only after the early decisions pass review ^[11]. Each is described in a few words. "We'd rather give someone one decision they can make as opposed to ten pages they will skim," Shopify wrote ^[16]. Anyone who has approved a large generated diff after reading its first file will recognise the problem. In my view the plan review is the best idea in Helix, because a bad slice order is cheapest to fix while it is still a short list.

Small slices also change what the model reads. A checkpoint fits in a small context window, so the agent reads the relevant part of the React Native code directly and skips the large spec file ^[12]. "The reference is the spec," the post says ^[12]. Shopify contrasts this with tools that turn gathered information into specs and task files, implement everything at once and hope the first result works ^[6].

The behavior gate runs on a command-line harness. Shopify's CLI exposes the same screen state and actions as the app. Its example is a home screen exposing analytics information and actions for navigating to other parts of the app ^[15]. The agent studies how the reference app works, replicates it and validates the result through CLI behavior tests ^[15]. A subagent writes those test cases from the reference code as integration tests from the user's perspective. Shopify says the tests are also used to push the agent toward edge cases outside the happy path ^[13].

The code reviews need a standard to hold code to. Shopify says it documented its opinionated architecture thoroughly so reviewers have one to enforce ^[17].

Shopify's stated reason for building guardrails at all is that LLMs are capable but hard to get consistent, maintainable results from out of the box ^[4]. The evidence that this design works at scale is still one company's account. The 12-week figure belongs to Shop ^[2]. The Shopify App, with more than 300 screens, is the migration now under way ^[3]. Shopify did not give a screen count for Shop or any Helix figures for gate failure rates or time per checkpoint, so the Shop timeline cannot be scaled into a schedule for the larger app.