Beyond Enthusiasm: A 5-Week, 86-Developer AI Coding Experiment

Wix Engineering ran a structured 5-week experiment with 86 developers across four countries to measure the real productivity impact of Claude Code versus Cursor and an internal tool called Astra. GitHub data (commits, PRs, lines changed) was analyzed using a Difference-in-Differences approach against 90-day baselines. Key findings: Cursor showed almost no measurable productivity lift despite users reporting feeling more productive; Claude Code showed significant output gains, but only after developers built trust and stopped micromanaging it. 47% of exit survey respondents said they accomplished tasks they wouldn't have attempted before. The experiment is now driving a company-wide rollout, but the author warns of an emerging 'AI Curse'—individual productivity exploding while organizational coordination, CI/CD pipelines, and review processes become the new bottleneck.