Shopify rebuilt 300 app screens in twelve weeks
Shopify rebuilt its merchant app with agents in twelve weeks, and CNCF measured how a firm's platform ownership shapes the AI infrastructure it ends up running.
How we Organise
CNCF and SlashData asked developers who owns their platform and how their firm runs AI workloads. The two answers moved together. Where individual teams manage their own tooling, 54 per cent report a hybrid approach to AI workflows; where the firm builds a custom internal developer platform, 5 per cent do. Of the developers at platform-building firms, 38 per cent run a separate dedicated AI platform and 33 per cent extend the platform they already have. Firms whose platform teams curate third-party tools sit between the two at 18 per cent hybrid, and a quarter of them rely mainly on a vendor’s AI platform.
We tell a client that how they split platform ownership is already a decision about their architecture, taken whether or not anybody writes it down.
Your systems will divide where your teams divide, whether you intended it or not.
Four ownership models produced four distinct shapes of AI infrastructure, which is what that claim predicts. Direction is the thing the report cannot settle. It publishes no sample size and no sampling frame. The relationship is an association. A firm that already runs fragmented infrastructure is plausibly the kind of firm that lets each team choose its own tools. CNCF is also reporting on the ecosystem it convenes.
How we Build
Shopify moved its merchant app off React Native to Swift and Kotlin. That is more than 300 screens, a watch app and widgets, rebuilt in twelve weeks by about thirty engineers working in pairs. Agents wrote most of the code. The team tried writing a specification per screen first, then abandoned it. A specification written for the first checkpoint was wrong by the third, and nobody wanted to maintain a description of behaviour the product already held. The old screen was the reference instead. An internal tool split each screen into small checkpoints and verified every one against the app still in production. Android launch time fell from 1.71 seconds to 0.87 at the median, and crash-free sessions went from 99.66 per cent to 99.98.
A position we hold is that a core gets replaced one use case at a time, with the old surface measurably smaller after each one.
Each use case carries a defined slice of core replacement, and the consumed surface is shrinking.
Shopify’s account describes that mechanism and reaches for no other. It does not settle the general case. The figures are Shopify’s own and nobody outside the firm has read them. No interface count is published, so the shrinking surface appears in the account of the checkpoints and not in a series. The whole method rests on having a working app to check each screen against. A firm replacing a core with nothing still running has nothing to check against. Two earlier posts from the same programme have already appeared here, the Reflex architecture and the mobile test-suite rebuild, so this is the third document Shopify has published about one piece of work.