AI coding agents are becoming more capable, but evaluating them is harder than it looks. Most benchmarks focus on a single dimension of agent capabilities; for instance, the popular SWE-Bench benchmark only ...