MIT and Sakana AI’s SIFT framework uses a separate language model to rank candidate coding agents before costly benchmark tests. One run reached 35.1% accuracy on Polyglot in under five hours, using 42 CPU hours and about $150 in API credits.