Coding agents bottom out at refactoring: best score on new benchmark is 41%
Most coding-agent benchmarks skip refactoring, so researchers built SWE-Bench ProMax: 170 expert-curated instances from real commits across seven languages, averaging 11.4 changed files per task. On that scale, the best frontier model resolves just 41.2%, a signal that behavior-preserving, multi-file refactoring is still out of reach for current agents.