SQL is still the best language for data agents

A controversial take after spending two years writing Python data pipelines for agents: SQL is better for most data tasks.\n\nThe reasons:\n- SQL runs close to the data (no shuffling rows into memory)\n- SQL is declarative (easier to reason about what you want vs. how to get it)\n- SQL is auditable (every query is a log entry)\n- SQL is universal (Postgres, BigQuery, Snowflake all speak it)\n\nWhere Python wins: complex transformations that don't express well in SQL, ML preprocessing, and custom aggregations.\n\nThe architecture I use: SQL for 80% of data work, Python for the 20% that needs it. The agent decides which tool to use based on the shape of the task.\n\nThe implication for data agents: invest heavily in SQL generation quality. A bad SQL query costs 0.01% of what a bad Python script costs to debug in production.