Nvidia's SoL-Pi cuts coding-agent token use by up to 49% by tuning the harness

The Decoder reports that Nvidia's SoL-Pi system cuts coding agents' token usage by up to 49 percent with little change in performance. The reported mechanism is optimization of the control layer — the harness — between the model and its environment, not the model itself. According to The Decoder, a research agent tested 152 approaches across more than 3,000 runs to develop the system. The same report notes the gains were smaller on other benchmarks, so the 49 percent figure should be read as a best case rather than a general result.