NVIDIA's SoL-Pi Cuts Coding Agent Tokens by 44%—Harness Engineering Is Now Its Own Field

NVIDIA Labs released SoL-Pi, a harness-level RSI (Recursive Self-Improvement) framework that runs auto-research loops across many diverse environments. On the 51-task EdgeBench benchmark, SoL-Pi delivered GPT-5.6 Sol and Opus 5 performance comparable to the Pi harness baseline while cutting recorded token traffic by 44.7%-49.0% and reducing API cost by roughly one third. GitHub stars reached 2.26k.

The four mechanisms that survived the auto-selection loop are: action execution, context compaction, observation handling, and delegated reading. They all sit on the hot path of token traffic, which is why the cumulative savings land close to half.

Why this is more than another NVIDIA paper

The paper also signals a quiet split between model engineering and scaffold engineering. The seven related works surfaced by Semantic Scholar (openJiuwen, StarHarness, HarnessDev, Prime Agent, RSIAgent, PILOT, Hierarchical Self-Improvement) all attack the same problem from different angles, suggesting that harness optimization is now a research domain on its own.

Code is open source at https://github.com/NVlabs/SoL-Pi. Early adopters running GPT-5.6 Sol or Opus 5 through Claude Code or Codex can expect roughly $8.75-13.50 in hourly savings, or $4.36-5.71 versus the Pi harness.

The catch: EdgeBench is NVIDIA's own eval set, and the gains only have public data for GPT-5.6 Sol and Opus 5. Whether open-weight coding models (Qwen3-Coder, GLM-4.6, Kimi K2.8) see the same 44% reduction is still an open empirical question.

Practical takeaway for agent infra teams: before swapping a smaller model, try swapping the harness first. The savings may be larger than the model delta.