AI agents that write and test code, not chat, master enterprise databases

Inside a big company, getting an answer out of the data is a relay race. The person who owns the data hands off to an engineer who structures it, who hands off to an analyst who finally queries it. Every handoff loses information and time.

A team built Data Intelligence Agents (DIA) to collapse that relay into one system of three AI workers. One reads raw data and figures out what it means. One designs the structure to hold it — the blueprint of tables and columns. One writes the database queries that pull answers out.

The key design choice: these agents don't just describe what to do in words. They actually write real code, run it, check whether the output is correct, and fix it when it breaks — the same loop a human programmer uses. They also share a common memory, so a trick that worked on one job can be reused on the next. A human expert reviews each piece before it's trusted.

The team tested the query-writing agent hard, on its own with no human help, across seven standard tests covering different question types and four different database languages. It matched or beat the best previously published scores on all seven.

What's untested in public is the other two agents — the reading and structuring ones — which the paper describes but doesn't benchmark. Their strength rests on the system already running for paying customers.

Why you should care: The same agent handled four database dialects and seven test types with only its plain-language instructions changed — no rebuild per task. That's what makes it deployable across messy real company data, not just one tidy benchmark.