Skip to content- Sources: primary
- Summary: About 30 engineers in pods of two rewrote the 300-screen Shopify app from React Native to native Swift and Kotlin in 12 weeks. Each engineer steered several agents, and a tool called Helix checked each small checkpoint against the old app. Agents give every PR an adversarial first review and fix merge conflicts and CI failures, but a human still reviews every PR. A self-healing loop runs every 15 minutes, reproduces new crashes and opens fix PRs. Android P50 launch time fell from 1.71s to 0.87s, and crash-free sessions rose from 99.66% to 99.98%.
- Why it matters: The post shows which parts of a team's process limit agent speed once code is cheap: build machines, review, merging and coordination.
- Background: Teams migrating with agents usually face a choice between writing specs and using the old system as the reference. Shopify tried spec-driven development first and dropped it because the specs went stale. It used the running React Native app as the spec instead. An agent turns each day's review comments into updated architecture rules and skills, so the number of requested changes per PR kept falling.