NVIDIA Open-Sources SoL-Pi, a Self-Evolving Agent Harness That Cuts Token Use Nearly in Half
NVIDIA's NVlabs published and open-sourced SoL-Pi, a self-evolving agent harness that recursively scales auto-research loops to cut token waste before agents are deployed across thousands of parallel instances. Official numbers show 44.7% less token traffic and roughly 50–54% lower API costs; both the GitHub repo and Hugging Face paper page are live.
Confirmed
- NVIDIA released the full technical report on Hugging Face and open-sourced the GitHub repo (NVlabs/SoL-Pi) (@nvidia, @aigclink).
- Method: inspired by RSI, it recursively extends auto-research loops at the harness layer, letting agents discover harness improvements across codebase- and verifier-driven environments, keeping only reusable mechanisms that survive selection (@omarsar0, @nvidia).
- Official numbers: 44.7% reduction in token traffic; 35–64% fewer tokens and 50–54% lower API cost vs. the native harness; 45–49% fewer tokens vs. base Pi (@aigclink, @akhaliq).
- Engineering: SoL-Pi is an independent extension of Pi (not an official distribution), packaging four reusable efficiency mechanisms installable on unmodified Pi, reducing repeated model turns and context replay (@aigclink).
Why it matters
- @omarsar0 notes the core contribution is letting the harness itself evolve under selection pressure rather than relying on manual tuning, offering a new paradigm for automated agent infrastructure optimization.
- At thousands of parallel unattended agent instances, per-instance token savings compound into significant cost gains (@songhanmit).
2026-09-18 ~ 2026-09-19 · 6 related posts
Primary sources
- NVIDIA's SoL-Pi Cuts Coding Agent Token Traffic ~49% and API Cost by a Third — nvidia ·
- NVIDIA's SoL-Pi auto-evolves agent harnesses, cutting tokens ~50% and API costs ~33% with no quality loss — omarsar0 ·
- NVIDIA open-sources SoL-Pi, an agent harness that cuts tokens 35–64% and API cost in half — aigclink ·
- [source] NVIDIA's SoL-Pi Cuts Coding Agent Token Traffic ~49% and API Cost by a Third — nvidia · 2026-09-18
- NVIDIA's SoL-Pi agent harness cuts token traffic ~47% and API cost by a third — _akhaliq · 2026-09-18
- [source] NVIDIA's SoL-Pi auto-evolves agent harnesses, cutting tokens ~50% and API costs ~33% with no quality loss — omarsar0 · 2026-09-19
- [source] NVIDIA open-sources SoL-Pi, an agent harness that cuts tokens 35–64% and API cost in half — aigclink · 2026-09-19
- NVIDIA's SoL-Pi GitHub repo: auto-research loops for efficient agent harnesses — aigclink · 2026-09-19
1 near-duplicate retellings: songhan_mit