Nvidia reduces coding agent tokens by almost half with SoL-Pi

Nvidia has published SoL-Pi, an approach that cuts down on coding agent tokens by nearly half through better harness optimization instead of model optimization. SoL-Pi doesn’t change the actual model, it changes the harness (i.e. instructions/policies) associated with the model. After testing 152 different configurations across almost three thousand trials, they discovered the best configuration cut down on tokens by 49% while only sacrificing 6.3% of quality (93.7%). On the EdgeBench benchmark, SoL-Pi cut down on tokens by 54.3% compared to Claude Code.

It looks like these improvements are legitimate and haven’t sacrificed too much in terms of answer quality. Also, there was less backtracking and extra actions performed by the agent. Remember, this is still a research paper from Nvidia so we can’t use SoL-Pi to optimize Claude Code just yet. This may not work quite as well for other types of problems like math olympiad questions, but I like how they focus on optimizing the harness around the model rather than the model itself since there seems to be a lot of unnecessary context being carried around by the agent.