A preprint benchmark finds coding agents struggle to locate code across large software systems, with the best setup completing 26.3% of tasks.