Small subsets of agent benchmarks for comparing systems at a fraction of the cost. Each card says how closely it tracks the full benchmark.