~24K lines of C. 960+ tests. Zero human code. Three AI agents—Challenger, Writer, Reviewer—collaborated in adversarial rounds until every test passed under AddressSanitizer. The result beats PostgreSQL on 16 of 19 batch benchmarks.
The three agents worked in iterative rounds. The Challenger would study the current
codebase and produce .sql test files exercising corner
cases—empty tables, NULL handling, multi-column ordering, stale index entries
after deletes, and more. The Reviewer would read the source and annotate it with
actionable comments flagging code-quality issues, missing edge-case handling,
and architectural improvements. The Writer would then run the new tests, address the
comments, diagnose failures, and ship fixes. This continued until the full suite passed
cleanly.
Write code, run 960+ tests, fix failures, repeat. The adversarial model drove correctness the same way rigorous code review does on a human team—except all three sides were machines.
The result: a recursive-descent parser, block-oriented vectorized executor, arena-based memory management, and PostgreSQL-compatible wire protocol—all without a single line of human-written C.
Architecture · Benchmarks vs PostgreSQL · Testing methodology · Try it in the browser