The impact of time(outs) on agentic benchmarks

When we launched WolfBench, a four-metric evaluation framework built on Terminal-Bench 2.0, we started with a fixed 2-hour timeout per task. After analyzing nearly 10,000 task results across 8 models,…