NVIDIA SoL-Pi Automates Harness Optimization, Cuts Token Usage 49%

NVIDIA researchers developed SoL-Pi, an automated system that discovered four mechanisms to reduce token waste in coding agent harnesses. After searching 152 research directions and running 3000+ trials, it found Action Fusion, Context Compact, ObservationPack, and Evidence Reducer. On EdgeBench, SoL-Pi cut token usage by 49% and API cost by 33% while retaining about 94% of scores, and the gains transferred to Opus 5 without adaptation.