How coding agents wrote Rust that beats state-of-the-art libraries
A benchmark-graded prompt loop pushed agents 4-15x past umap-learn; the prompt kit is open, but HN warns the harness does the work, not the model.
Copy markdown
The recipe: grade every attempt against a real library
Max Woolf didn't tell the agent to "make it faster" — he set a hard bar (every criterion benchmark ≥1.2x faster), banned unsafe code and fake parallel runs, and pitted models against named crates. The payoff: a Rust UMAP 4–15x faster than the umap-learn Python bindings and 2–4x faster than umap-rs, plus a templating engine that beat minijinja and tera.
The kit is open — you can rerun the loop this weekend
The "Ur-Prompt" and a Rust AGENTS.md are posted as public gists, anti-cheating rules included (no disabled tests, independent verification, read-only benchmarks). Point Claude or Codex at your own hot path, let it iterate to convergence, then hit it with a "try a fundamental breakthrough" pass for another 1.2–1.5x.
The catch: the feedback loop does the work, not the model
Performance engineers on HN pushed back: LLMs reason poorly about cache and low-level instructions, and agents will "benchmaxx" to the test shape if you let them. The gains came from the measurement harness and guardrails — without a real benchmark to grade against, the same prompts flail. Validate against different input shapes before you trust the numbers.