NVIDIA open-sources SoL-Pi, an agent harness that cuts tokens 35–64% and API cost in half
aigclink · x · 2026-09-19
NVIDIA's NVlabs open-sourced SoL-Pi, a token-efficiency-focused agent harness: 35–64% fewer tokens and 50–54% lower API cost vs the native harness, and 45–49% fewer tokens with roughly 1/3 the cost vs base Pi — without degrading task completion.
The idea: AI automatically audits AI workflows to find waste (repeated replays, oversized content transfers, unnecessary thinking turns, expensive models doing cheap work), then restructures the harness — making AI efficient before scaling token-hungry auto-research (RSI) loops.
Key mechanisms:
- Merging high-certainty "edit-then-test/build" steps into single actions
- Compression only when expected savings repay the rewrite cost; large content stored in a repo with context holding just a "pickup code" and short excerpts
- Long-log reading outsourced to cheap models, whose receipts must be verified line-by-line against archived logs before expensive models proceed
Iron rule: savings must come from eliminating waste, never from doing less work. The GitHub repo has 2.3k stars.
More from coding & agent
- Orgo, a YC-backed platform giving AI agents their own cloud desktops, claims 1,000+ teams — Scobleizer · 2026-11-04
- Founder: Vibe-coded a PMF-successful product, then hired pro devs to fix the mess — sujingshen · 2026-11-04
- One prompt to audit cross-region traffic: dev cut trans-Atlantic latency by moving a service — DanielLockyer · 2026-11-04
- Agent silently completed only a quarter of a 200-module build and stalled waiting for confirmation — Promptmethus · 2026-11-04
- ChatGPT Agent mode can scan LinkedIn and Indeed and build a tailored job-application spreadsheet — nikola_mr64990 · 2026-11-04
- Blogger auto-applied to 500 jobs via ChatGPT and got 12 interview calls in 24 hours — nikola_mr64990 · 2026-11-04