agentsfactbullishPerformance on expensive agentic benchmarks can be accurately predicted by performance on a small, carefully selected subset of atomic evaluation instancesComputation and Language27 Jul 2026http://arxiv.org/abs/2607.02032v1