Nvidia says tuning the harness, not the model, can nearly halve coding-agent token use
Automated harness optimization moves agent cost-cutting off the model, but gains concentrate on the authors' chosen benchmark.
Original event 2026-09-26
Nvidia researchers report in a preprint that SoL-Pi, a system that uses a research agent to rewrite a coding agent's harness, cuts token use by 49 percent in its leanest variant while retaining 93.7 percent of the Pi harness's score.
The system explored 152 directions across 535 executable environments and more than 3,000 runs, producing four mechanisms including action fusion and context compaction. At current API prices, the authors estimate savings of $8.75 to $13.50 per hour versus native Codex and Claude Code harnesses.
The gains do not hold everywhere: on 63 Terminal-Bench 4 tasks, SoL-Pi solved 15 while Codex and Pi each solved 18. The authors themselves note that automatically optimized harnesses tend to overfit their training tasks.