Data Intelligence Agents system beats 7 SQL benchmarks with autonomous coding agents

Data integration in enterprises is often bottlenecked by repeated handoffs between data owners, engineers, and analysts. A new paper on arXiv presents Data Intelligence Agents (DIA), a multi-agent system designed to streamline this workflow. DIA consists of three specialized agents: a Data Interpreter that understands raw data, a Schema Creator that structures it, and a Query Generator that writes and validates SQL queries. Unlike traditional LLM-based approaches that emit text, DIA treats autonomous coding agents (ACAs) as a first-class abstraction. These agents generate, execute, validate, and repair concrete artifacts (e.g., SQL queries), and they leverage a shared memory for experience reuse. Domain experts review and approve each artifact before it moves to the next step.

In production for enterprise customers, DIA’s Query Generator was evaluated in fully autonomous mode across seven SQL benchmarks spanning four task categories (e.g., text-to-SQL, schema inference) and four dialects (PostgreSQL, MySQL, Snowflake, BigQuery). It matches or surpasses the best published results on all seven benchmarks. The key insight: an architecture grounded in execution—rather than pure text generation—generalizes across the data intelligence workload, with adaptation confined to natural-language instructions. This approach reduces human overhead and accelerates data pipeline creation.