Cursor's rebuilt agent swarm reconstructs SQLite from its manual in four hours

Cursor said on July 20, 2026 that its rebuilt agent swarm reconstructed the SQLite engine in Rust from its 835-page manual alone, reaching an 80 percent pass rate in four hours, in a research post.

An agent swarm is a group of AI coding agents that work in parallel under a coordinator rather than one model working alone. Cursor ran the test with no source code, test suite, binary or internet access, grading the output against the sqllogictest suite. Using xAI’s Grok 4.5, the new swarm hit 80 percent in four hours; the old system was paused before the two-hour mark, and every new-system configuration reached 100 percent test completion versus 11 to 77 percent before.

The efficiency gains are the sharper story. Code volume for one model pairing fell from 64,305 lines to 9,908, and merge conflicts across a four-hour run dropped from more than 70,000 to under 1,000. Total cost ranged from $1,339 for one hybrid configuration to $10,565 for a single-model GPT-5.5 run, with worker agents accounting for 69 to 90 percent of token volume.

Cursor published the resulting Rust codebase as minisqlite on GitHub.

The results are Cursor’s own, run on a single well-specified task, and a manual-only reconstruction of a documented system is not the same as building novel software from scratch. The company has not released independent verification of the pass rates or costs.

Whether the coordination gains hold on messier, less-documented codebases, or degrade outside the controlled SQLite setup, is what the next round of tests will show.