Shopify Helix blocks coding agents until evidence passes review

AI Labs reconstructs Shopify’s Helix pattern as an orchestrated loop: split a large change into ordered checkpoints, then require behavior tests, visual comparison, adversarial code review, and a final human decision before work advances. Fresh subagents isolate each task, while failed gates send the work back for another attempt. Why it matters: Long agent runs can hide compounding mistakes inside one large diff. Small checkpoints make early decisions cheaper to correct, and blocking gates turn quality requirements from prompt advice into conditions the workflow must prove before it continues.

Try this: Test one feature with three ordered checkpoints. Require each diff to show passing tests, a visual comparison where relevant, and an adversarial review before the agent starts the next checkpoint.