“A Databricks Research benchmark found that the best coding agents could answer novel data questions correctly about 50% of the time.”