SoL-Pi paper: auto-research loops save $4-13 per hour on coding agents
alex_verem · x · 2026-09-23
arXiv details for SoL-Pi (2609.20519, submitted Sep 17, 2026): 14 authors including Haozhe Liu, Song Han and Enze Xie; 15 pages, code and project page released.
- As coding agents move to unattended exploration, token efficiency becomes critical for recursive self-improvement; the paper scales auto-research loops at the harness layer across increasingly diverse environments.
- Four selected mechanisms span action execution, context compaction, observation handling, and delegated reading, forming SoL-Pi with improvements that transfer beyond the development setting.
- Numbers: on EdgeBench across GPT-5.6 Sol and Opus 5, performance matches Pi while recorded token traffic drops 44.7–49.0% and API cost falls 1/3 — estimated hourly savings of $8.75–$13.50 vs native Codex/Claude Code harnesses and $4.36–$5.71 vs Pi.
Related event: NVIDIA's SoL-Pi Cuts Agent Token Costs by Nearly Half(3 posts)→
More from coding & agent
- Google launches Gemini 4 Carbon with frontier cybersecurity and long-horizon skills — SRSchmidgall · 2026-11-11
- Turning 1,306 X bookmarks into a searchable wiki with a Grok Bot — Scobleizer · 2026-11-11
- DeepSeek Harness sparks agent platform debate: custom FDE vs selling seats — teortaxesTex · 2026-11-11
- Jason Lemkin's list: If you can't ship in week one, don't join an early-stage startup — alvelda · 2026-11-11
- CMU AI Agents course releases full lecture on computer use agents (CUAs) — kohjingyu · 2026-11-11
- Tailscale turns Aperture into a team-wide AI skill hub with Skills over MCP distribution — evielync · 2026-11-11