“In an internal Databricks benchmark of real employee questions, the best coding agents were able to solve the problems correctly about 50% of the time.”