NVIDIA · 2026-09-17 · major
SoL-Pi — NVIDIA's harness extension cuts coding-agent tokens by about half
SoL-Pi is an MIT-licensed extension for the Pi coding agent from NVIDIA's research lab. It packages four token-saving mechanisms found by an automated research loop, cutting token use 45-49% while keeping about 94% of Pi's EdgeBench score.
Four efficiency tricks, picked by an automated research loop out of 152 candidates, packaged as a drop-in Pi extension.
Key specs
| Token reduction | 45-49% |
|---|---|
| Score retained | ~94% |
Quick facts
| Maker | NVIDIA (NVlabs) |
|---|---|
| What it is | Extension for the Pi coding agent |
| License | MIT |
| Requires | Node.js 22.19+, pi-coding-agent 0.85.1 |
| Evaluated on | EdgeBench, 51 tasks |
| Mechanisms | Action Fusion, ObservationPack, Evidence-Preserving Reducer, Online Context Compact |
| Paper | arXiv 2609.20519 |
What is it?
Four reusable efficiency mechanisms arrive in SoL-Pi, a standalone extension that installs into the Pi coding agent. Action Fusion merges a file edit and the command that follows it into one tool call. ObservationPack archives big tool outputs on disk and leaves a handle plus a short excerpt in context. An Evidence-Preserving Reducer hands log-reading to cheaper agents, and Online Context Compact chooses when to rewrite context.
How does it work?
The four mechanisms were not hand-designed. NVIDIA's lab ran an auto-research loop that fanned 152 proposals across independent lineages; each one generated trajectory rollouts, analysed them with map-reduce, implemented a candidate mechanism, passed a separate reviewer, and then faced a fully isolated held-out set. Only 4 proposals survived. No agent inside the loop ever saw the held-out results, so the surviving mechanisms had to generalise rather than fit the training trajectories.
Why does it matter?
Token traffic is the bill for running coding agents, and SoL-Pi attacks it at the harness layer rather than by swapping to a weaker model. On EdgeBench it costs about a third less per hour than plain Pi and 50-54% less than the native Codex and Claude Code harnesses, while still returning roughly 94% of Pi's average score. Teams running agents continuously get most of the capability for half the traffic.
Who is it for?
teams running coding agents at scale
Frequently asked questions
- Does SoL-Pi work with Claude Code or Codex?
- SoL-Pi installs into the Pi coding agent specifically, not into Claude Code or Codex. Those two appear in the paper as the native harnesses SoL-Pi is measured against, where it reports 50-54% lower API cost. The mechanisms are described as reusable, but the shipped package targets Pi 0.85.1 and its extension interface.
- How much money does SoL-Pi actually save per hour?
- NVIDIA's lab reports SoL-Pi saving $4.36-$5.71 per hour compared with running plain Pi, and $8.75-$13.50 per hour compared with the native Codex and Claude Code harnesses. Those figures come from the 51-task EdgeBench run, so real savings depend on how closely your workload resembles long-horizon coding tasks.
- Can I turn individual SoL-Pi mechanisms off?
- Yes. SoL-Pi reads a sol-pi.json config from either .pi/sol-pi.json in a project or ~/.pi/agent/sol-pi.json for a user. Each of the four mechanisms is its own boolean. The repo suggests a conservative starting configuration that enables Action Fusion and ObservationPack while leaving the Evidence-Preserving Reducer and Online Context Compact switched off.
- What does SoL-Pi give up in exchange for fewer tokens?
- About 6% of score. SoL-Pi retains roughly 94% of Pi's average EdgeBench result on both model backends it was tested against, so the trade is a small accuracy drop for 45-49% fewer tokens. Whether that is worth it depends on how often a task failure costs more than the tokens saved.
Try it
pi install git:github.com/NVlabs/SoL-Pi