Researchers introduce Mid-Harness, a method that allocates test-time compute at the model-harness boundary to improve action reliability and trajectory success in terminal agents. They evaluate Mid-Harness on various models and benchmarks, finding that it can improve performance and reduce estimated token cost.