As coding agents shift from supervised code completion to unattended, round‑the‑clock exploration, their work expands from isolated predictions to long trajectories of reasoning, tool use, and feedback. Consequently, token efficiency becomes crucial for scaling recursive self‑improvement. We adopt an RSI‑inspired approach at the harness layer, scaling auto‑research loops across increasingly numerous and diverse environments for harness roll‑outs. At this scale the process yields reusable improvements that transfer beyond the development setting, moving automated harness discovery toward production‑level outcomes. Four mechanisms survive selection and constitute SoL‑Pi:
- Action Execution
- Context Compaction
- Observation Handling
- Delegated Reading
On the 51‑task EdgeBench benchmark, SoL‑Pi achieves performance comparable to Pi built on GPT‑5.6 Sol and Opus 5, while reducing recorded token traffic by 44.7%–49.0% and cutting API cost by roughly one third. In hourly terms, SoL‑Pi saves $8.75–$13.50 relative to native Codex and Claude Code harnesses, and $4.36–$5.71 relative to Pi.
Review: SoL‑Pi demonstrates a practical path to simultaneously reduce token usage and cost through recursively scaled auto‑research loops, offering a solid blueprint for next‑generation efficient coding agents.