Needle: The benchmark your search engine can't memorize

Needle is a live, open-source benchmark designed to evaluate search engines on agentic traffic, using queries from real agent logs and synthetic intents that reflect production usage. It runs continuously, publishes all queries and metrics publicly, and provides full evaluation code on GitHub to enable reproducible testing. The benchmark addresses overfitting and data-leakage issues that undermine static tests like BrowseComp.