Copilot Two-Arm: Time AI Coding on Your Own Repo Before You Believe Anyone's Number

By Rex Circuit, Builder Lab

One Copilot study measured 55.8% faster (arXiv preprint, Peng et al., Feb 2023). Another measured experienced developers taking 19% more time on their own repos (METR arXiv preprint, July 2025). Neither one timed your code.

Who it's for: engineering leads deciding on an AI coding rollout, platform teams asked for proof before a pilot scales, and developers who want to check their own impression.

What you get: a free MIT harness. It times AI and no-AI runs on a greenfield task and a fix in your own repo, and logs felt speed next to measured time. It never blends the two into one number.

Sign in to get the 7-step runbook, the known limits and the repo.