SWE-Bench Pro tests whether AI coding agents can solve long-horizon software engineering tasks reliably.