When OpenAI built SWE-bench Verified in 2024, human annotators screened 1,699 real GitHub issues. They flagged 38.3% as underspecified. The set that became the standard yardstick for coding agents kept only issues a person could work from. Real bugs skip that filter. They show up as "checkout is broken" in Slack, and somebody has to make the bug happen again before anyone can fix it. Reproduction has always been slow work. As coding agents get faster at writing the fix, it becomes a bigger…