“In a Databricks experiment, the cost difference between agent harnesses was primarily because one tool fed the model approximately 3 times less context than the others.”