TL;DR: A team of agents rebuilt SQLite from its manual in Rust, passing all tests, with cost varying 15x by model mix.
Summary: Cursor AI used a team of agents to rebuild SQLite from its 835-page manual into a Rust replica that passed 100% of a held-out test suite. The cost varied 15x depending on the mix of models used for the task.
Why it matters: This shows that large codebases can be accurately recreated from documentation alone, reducing manual coding effort. AI builders should experiment with multi-model strategies to balance cost and performance for complex code generation tasks.
Source: x_com