DigestAI news desk

Cut through the AI noise.

Agents & Tools4 min read

Nvidia's SoL-Pi cuts coding agent token usage by nearly half

Nvidia researchers released a paper describing SoL‑Pi, a system that automatically optimizes the harness of coding agents. The harness sits between a model and its environment and controls how an agent sees states, runs actions, and processes feedback. According to the researchers, SoL‑Pi cuts token usage by almost half while keeping performance roughly the same.

1 source

Key points

  • SoL‑Pi reduces coding agent token usage by 49 percent while maintaining 93.7 percent of Pi harness score.
  • The system explored 152 directions across 535 environments, generating over 60,000 agent‑environment interactions.
  • Four mechanisms—Action Fusion, Online Context Compact, ObservationPack, Evidence‑Preserving Reducer—each cut token usage by merging steps, trimming context, summarizing outputs, or routing logs to cheaper models.

The study explored 152 directions across 535 executable environments, including 495 tasks from GitHub issue‑pull‑request pairs and 40 synthetic test cases. It generated more than 3,000 runs and over 60,000 agent‑environment interactions. Four mechanisms—Action Fusion, Online Context Compact, ObservationPack, and Evidence‑Preserving Reducer—each reduce token usage by merging steps, trimming context, summarizing outputs, or routing logs to cheaper models.

On EdgeBench’s 51 public tasks, the most efficient variant of SoL‑Pi used 49 percent fewer tokens and achieved 93.7 percent of Pi harness score. Users prioritizing performance can beat Pi’s score by 5.3 percent while still saving tokens. In dollar terms, the researchers estimate savings of $8.75 to $13.50 per hour versus native Codex and Claude Code harnesses, and $4.36 to $5.71 per hour versus Pi.

The story so far

2 episodes →
  1. Nvidia's SoL-Pi cuts coding agent token usage by nearly halfthis story
Full story from The Decoder · by Jonathan KemperOpen source ↗

Nvidia's SoL-Pi system cuts coding agent token usage nearly in half by optimizing the harness

The Decoder · 26 September 2026

Loading the full article…

This text was published by The Decoder and written by Jonathan Kemper. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗

Topics · follow one to build your own front page

The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.

Comments

via GitHub Discussions

More in Agents & Tools

All →

Related stories