Shopify says AI agents drive cleaner code, not just faster code

The narrative around AI coding tools has been stuck on speed — how many lines can Copilot spit out, how fast can Cursor autocomplete a function. But Shopify’s recent experience suggests a more interesting dynamic: AI agents are actually pushing engineering teams toward better architecture, not just faster output.

Speaking at a developer conference, Shopify engineering leaders described how their internal AI coding assistants perform demonstrably worse on messy, tightly-coupled codebases. When an agent encounters a function that does twelve things, touches five services, and has no clear contract, it stumbles. It hallucinates. It breaks things. The fix, it turns out, is the same thing senior engineers have been arguing for years: write smaller, single-responsibility functions with explicit interfaces and comprehensive tests.

“Agents just want the same things as humans: easily-readable code, explicit contracts, and helpful feedback,” one Shopify engineer noted. The company has been on a multi-year journey to simplify its massive Rails monolith, and AI tooling has become an unexpected forcing function. Code that’s hard for an LLM to reason about is also hard for the new hire who joined last week. The feedback loop is tighter now — if your pull request confuses the AI reviewer, it’s probably confusing your teammates too.

This cuts against the prevailing fear that AI will enable a flood of sloppy, generated code that nobody understands. In Shopify’s case, the opposite is happening. Teams are refactoring legacy modules not because some architecture document told them to, but because their AI tools keep producing garbage when pointed at those modules. The incentive to clean house is immediate and practical, not theoretical. Other engineering orgs watching this will likely reach the same conclusion: if you want AI to be useful, your codebase needs to be legible first.