Hugging Face has published a research paper titled SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness, which examines how coding agents can be made more token-efficient as they shift from supervised code completion toward unattended, around-the-clock exploration.
The authors note that as agents take on longer trajectories of reasoning, tool use, and feedback, token efficiency becomes a key constraint on scaling recursive self-improvement. Their approach applies an RSI-inspired method at the harness layer, scaling auto-research loops across increasingly numerous and diverse environments for harness rollouts.
According to the paper, this scaled process produces reusable improvements that transfer beyond their original development setting, moving automated harness discovery closer to production-level outcomes. Four mechanisms survived the selection process and form SoL-Pi, covering action execution, context compaction, observation handling, and delegated reading.
On the 51-task EdgeBench evaluation, SoL-Pi reportedly achieves performance comparable to Pi across GPT-5.6 Sol and Opus 5, while reducing recorded token traffic by 44.7–49.0% and API cost by roughly one third. The paper estimates hourly savings of $8.75–$13.50 relative to native Codex and Claude Code harnesses, and $4.36–$5.71 relative to Pi.