Nvidia SoL-Pi cuts coding agent token use by optimizing the harness

Nvidia researchers describe SoL-Pi, which optimizes the harness between coding models and their environment, cutting token use by up to 49% with little performance change on their main test; gains were smaller on other benchmarks. A research agent tried 152 approaches over more than 3,000 runs to automate harness tuning instead of hand-editing traces.