Nvidia researchers report that SoL-Pi, a system that automatically optimizes the harness controlling coding agents, cuts token usage by 44.7 to 49 percent while maintaining performance comparable to the original Pi harness on EdgeBench. The search produced four mechanisms, including merged action steps and context trimming. The approach applies to harnesses used by systems like Codex, Claude Code, and OpenClaw, with estimated savings of $8.75 to $13.50 per hour versus native Codex and Claude Code setups.
No score is assigned. Sources and their independence are shown in the citation chain below.