Nvidia's SoL-Pi system cuts coding agent token usage nearly in half by optimizing the harness
decoder ·
Nvidia researchers have introduced SoL-Pi, a system that optimizes the control layer of coding agents to significantly reduce token consumption. By automating the analysis of agent traces and proposing harness modifications, the system cuts token usage by nearly half while maintaining comparable performance levels. The approach explores over 152 directions across 535 environments, identifying four key mechanisms such as Action Fusion and Online Context Compact to eliminate redundant work. On the EdgeBench benchmark, the most efficient variant reduced costs by 49 percent and achieved 93.7 percent of the original Pi harness's score. While effective on specific benchmarks like EdgeBench, results were mixed on other tests like Terminal-Bench 4, where it solved fewer tasks than competitors. The study highlights that optimizing the harness can yield substantial savings compared to native implementations of Codex and Claude Code. However, the authors note that these efficiency gains may come with trade-offs, such as reduced prompt cache reuse. This development addresses the growing cost burden of agentic AI, which has seen token usage increase fourteenfold since early 2026.
decoder