Vivek Raghunathan, SVP of engineering at Snowflake, outlines a five-stage maturity model for AI coding tool rollouts that most organizations compress or skip, leading to uneven outcomes despite high adoption. The sequence begins with "let chaos reign," where the only metric is broad usage across diverse tasks, not output measures like lines of code or PRs merged. Stage two requires actively discovering effective practices from "fearless explorers"; Raghunathan's team identified roughly 14 named patterns, such as planning in writing before coding, fencing agents into separate git worktrees to prevent collisions, and keeping orchestrating agents context-light for delegation. Stage three codifies these into a shared pattern language, modeled on Gang of Four design patterns, enabling concrete teaching and measurement. Stage four scales adoption via structural mechanisms—not mandates—such as regular "focus weeks" where the org dedicates time to learn the paved path or pioneer new techniques. Only after these stages does stage five, reorganizing roles and org charts around verified AI-assisted workflows, become viable. Skipping the middle stages leaves a gap between engineers who transform their workflow and those using AI as advanced autocomplete. Raghunathan emphasizes that buying licenses is a one-time decision, whereas building a practice of discovery, codification, and deliberate spread is an ongoing management discipline that cannot be shortcut by premature restructuring or metric-chasing.
Read original
stackoverflow/blog