SoL-Pi cuts coding-agent token traffic 44.7-49% on EdgeBench

A coding-agent harness called SoL-Pi trims recorded token traffic by 44.7 to 49.0 percent on the 51-task EdgeBench evaluation while matching the performance of the Pi baseline it extends, according to a preprint on arXiv. API cost falls by "about one third," with hourly savings estimated at $8.75 to $13.50 against native Codex and Claude Code harnesses, and $4.36 to $5.71 against Pi.

The framing is operational rather than model-centric. "As coding agents move from supervised code completion to unattended, around-the-clock exploration, their work expands from isolated predictions into long trajectories of reasoning, tool use, and feedback," the authors write. Token efficiency at the harness layer, in their telling, becomes the scaling constraint for recursive self-improvement, not the underlying model.

Four mechanisms survive the authors' selection process and make up the system: "action execution, context compaction, observation handling, and delegated reading." Testing runs across GPT-5.6 Sol and Opus 5.

The abstract does not publish per-mechanism attribution, per-task accuracy, or the absolute pass rate that "comparable to Pi" is anchored against.