The paper argues that the infrastructure surrounding language model agents, known as the harness, significantly influences performance more than the models themselves. It calls for transparency in harness specifications to ensure accurate evaluations of long-horizon agent capabilities.