Cursor had a team of agents rebuild SQLite from its 835-page manual

What's this about?

Cursor pointed a team of coding agents at the 835-page SQLite documentation and had them re-implement the whole database engine in Rust — no peeking at SQLite's source. The replica passed 100% of a held-out test suite. Notably, total cost varied 15x depending on which mix of models did the work. Cursor also doubled usage limits for all individual and team plans.

What exactly is SQLite, and why is rebuilding it a big deal?

SQLite is the world's most deployed piece of software — the small database inside every phone, browser, and car. It's famously battle-tested and subtle (file formats, transactions, edge cases). Rebuilding it from *spec* — the written manual, not the code — is the classic "clean-room reimplementation" that used to take an expert team years.

What's the community saying?

1. The 15x cost spread across model mixes echoes the day's routing theme: which models you use, and for what, now matters more than which single model is "best."
2. Skeptics ask how much SQLite knowledge was already in the models' training data — is this spec-reading or memory recall?
3. Passing a held-out test suite ≠ production-ready: concurrency, corruption recovery, and 20 years of fuzzing aren't in the test suite.