Replay tasks catch coding-agent friction that benchmarks miss

Coding-agent adoption should check repository reading, scope control, test behavior, and recovery, not only public scores.

A coding agent can look strong on public comparisons and still be expensive to supervise in a real repo. A replay task should check whether the agent reads files first, follows existing patterns, edits the smallest necessary area, runs relevant tests, and recovers from failed commands. Those behaviors determine review burden. This is why I like replay sets as an adoption gate rather than a replacement for public benchmarks.