If you can’t score an agent, you can’t safely upgrade it. Evals are the unit tests of autonomous systems.