Today we’re open-sourcing hyper-𝜏-bench, a new long horizon agent evaluation that measures how well models can not only act as an agent, but construct one.