When optimizing software with coding agents, it’s easy to mistake noisy wall-clock improvements for meaningful gains. I’ve started using a proxy-first approach that measures the mechanism an optimization actually changes.