Agents rebuilt SQLite in Rust from the manual alone

Cursor reports that teams of coding agents built Rust database implementations from SQLite's manual and eventually passed every query in a withheld test suite. Its research report says the agents had no access to SQLite's source code, binary, tests or the internet. The suite checks whether database queries return expected results, so passing it establishes performance on those tests rather than complete interchangeability with SQLite.

The full experiment also separates agents that plan work from agents that implement it. Four controlled model configurations eventually passed the whole test suite at substantially different costs. This supports testing which models handle each role efficiently. The post's wider 15-fold cost comparison includes informally graded runs, which cannot establish equivalent quality across that entire range.