“A Databricks experiment showed a more than 2x difference in cost per task with the same quality when running the same model through different agent harnesses.”