“Human experts outperformed AI agents on the RE-Bench benchmark when given longer time budgets of 8 and 32 hours.”