Nvidia researchers developed SoL-Pi, a system that optimizes the control layer of coding agents to reduce token consumption. By automating the refinement of the harness—the interface between the model and its environment—the system achieves token savings of 44.7 to 49 percent while maintaining performance levels comparable to existing solutions.
The system utilizes a research agent that analyzes execution traces to propose and test improvements. During development, it explored 152 directions across 535 environments, generating over 3,000 runs and 60,000 interactions. To prevent overfitting, researchers strictly separated the search process from the EdgeBench evaluation benchmark.
SoL-Pi introduces four optimization mechanisms: Action Fusion, which merges consecutive steps; Online Context Compact, which trims context; ObservationPack, which summarizes tool outputs; and the Evidence-Preserving Reducer, which uses a cheaper model to process logs. These tools can be combined or used individually depending on performance needs.
Financial analysis indicates hourly savings of $8.75 to $13.50 compared to native Codex and Claude Code harnesses, and $4.36 to $5.71 compared to the Pi harness. In specific tests, such as a kernel optimization experiment, a swarm of 20 SoL-Pi workers reduced costs by 26.8 percent compared to a standard Pi-based swarm.
The researchers noted that while SoL-Pi is effective, results vary across different benchmarks. On Terminal-Bench 4, the system solved 15 out of 63 tasks, compared to 18 solved by Codex and Pi. Future research aims to explore pretraining harnesses across diverse tasks to further enhance efficiency and scalability.
