Nvidia's SoL‑Pi cuts coding‑agent token use nearly in half

Nvidia said a new system called SoL-Pi could cut coding agents' token usage by up to 49%, while keeping most of their original performance.

The system tests changes to an agent's "harness" — the code managing context, tool calls and feedback — rather than retraining the model, Nvidia said. A research agent proposed and tested 152 different approaches across more than 3,000 runs and 60,000 agent-environment interactions to find ones that held up.

Four techniques survived testing, Nvidia said: merging consecutive agent steps, trimming accumulated context, archiving old outputs with summaries, and distilling logs through a cheaper model. The best version kept 93.7% of the original performance on Nvidia's held-out benchmark, which could cut costs by $8.75 to $13.50 an hour versus unoptimized alternatives, according to the company.

The gains were smaller on other benchmarks, Nvidia said, and dropped noticeably on a set of terminal-based tasks. The harness was optimized only on transcripts from OpenAI's GPT-5.6 Sol model. Nvidia said it is not yet clear how well the same tweaks generalize to agents built on other models.