The arXiv study of Microsoft’s early-2026 rollout of Claude Code and GitHub Copilot CLI is the first large-N number I’ve seen that isn’t a vendor benchmark: tens of thousands of engineers, four months, merged PRs as the output proxy. Adopters shipped about 24% more merged pull requests than they otherwise would have, and the lift held — no novelty cliff.

The result everyone will quote is the 24%. The result that actually changes how you run a rollout is the other two: first use spread through social ties, not mandates or enablement decks, and sustained usage tracked how much someone already codes, not their tenure or team. That reframes agent adoption as a diffusion problem, not a training problem. If you’re rolling agentic tooling across an enterprise, the lever isn’t a better onboarding wiki — it’s making high-activity engineers visibly successful so the network does the propagation.

I’d push on the proxy, though. Merged PRs measure throughput, not the thing I care about operating AI in production: review load, defect escape rate, and the token bill — the paper itself flags annual spend in the millions. A 24% PR lift paired with a jump in reviewer time or a rising revert rate is a very different trade than it looks. The HN thread has engineers poking at exactly this.

Still, “peer adoption beats mandates and the effect persists” is a cleaner finding than most internal decks produce. If your org is measuring agent ROI purely on merged volume, what’s your counter-metric for the review debt it quietly moves downstream?