NVIDIA open-sources SoL-Pi: an AI-improved harness that cuts agent token use by up to 64%
新智元 · wechat · 2026-09-11
NVIDIA (Han Song's Efficient AI team) open-sourced SoL-Pi, an efficiency layer built on Pi's harness — installable with one line: piinstall git:github.com/NVlabs/SoL-Pi. Versus stock Codex / Claude Code harnesses it uses 35-64% fewer tokens, cuts priced cost 50-54%, and retains 94% of average score; in research workloads that's $8.75-13.5 saved per hour.
The mechanism came from an "AI researches AI" pipeline: 535 verifiable environments, 152 AI-proposed optimization ideas, then filtering via trajectory-based payoff estimates, AI-authored experiments with reviewer agents, and held-out verification — only 4 survived (1 in 40):
- Action Fusion: merges an edit with its obvious follow-up command into one local execution sequence, eliminating an intermediate model decision round.
- Online context compaction: evaluates at each completed subtask and only rewrites context when projected savings exceed the rewrite cost (including KV-cache invalidation).
- Observation Pack: archives large tool outputs to disk, keeps a short handle plus excerpt in context, and pages content back on demand.
- Evidence-Preserving Reducer: a cheap model summarizes long logs into a compact diagnostic receipt, which is verified line-by-line against archived logs before reaching the frontier model — blocking hallucination propagation.
Long-term, the team is betting on an RSI flywheel: closed-loop pretraining of the harness (agents collect tasks and build environments themselves) and compounding "efficiency for efficiency" — while model scaling laws are still debated, the harness scaling race has begun.
Related event: NVIDIA Open-Sources SoL-Pi, Cutting Agent Token Costs by Up to 64%(3 posts)→
More from coding & agent
- Ollama lets you customize ChatGPT's model selector with local and cloud models — ollama · 2026-09-11
- ChatGPT Desktop Codex app can now use local models via Ollama 0.34 — ollama · 2026-09-11
- Inspect evals: default continuation prompt can make agents misread bad actions as approved — AdtRaghunathan · 2026-09-11
- Claude Code creator: AI-written production code should be held to a higher bar — bcherny · 2026-09-11
- Dev culture gap: some burn a week of Codex quota daily, others never tried it — yihui_indie · 2026-09-11
- Modern Context Stack: a satire site skewering the data industry's tool-buying habit — juansequeda · 2026-09-11