NVIDIA Introduces SoL-Pi: Automatic AI Agent Harness Optimization

NVIDIA introduced SoL-Pi, a system that automatically improves AI agent harnesses and cuts token usage by nearly half without noticeable quality loss.

Instead of manual tuning, researchers run cycles of automatic experiments on the harness itself across different repositories and environments with verifiable outcomes. Only mechanisms that consistently pass selection make it into the final version.

Four mechanisms remained:

• action merging;
• context compression during execution;
• more efficient observation processing;
• reduced delegated reading while preserving important evidence.

On EdgeBench with 51 tasks, SoL-Pi matched the quality of the baseline system with GPT-5.6 Sol and Opus 5 while reducing token traffic by almost 50%.

According to the authors, this cuts API costs by about one-third. The savings are approximately $8.75–13.50 per hour compared with standard Codex and Claude Code systems and $4.36–5.71 per hour compared with the base harness.

The code is published on GitHub in NVlabs. The paper is available on arXiv.