AutoHarnessBench

Measures frontier models' capability to design and optimize their agent harnesses for a given task. Each model starts with a minimal harness, a task suite, and a fixed compute budget, then iteratively optimizes that harness to maximize held-out task performance.