NVIDIA Labs introduced SoL-Pi (arXiv:2609.20519, September 17, 2026) — an open system where coding agents themselves searched for improvements to their own harness: out of 152 proposed hypotheses, only four mechanisms passed honest selection, reducing token traffic by 44.7–49.0% and API costs by about a third on the isolated EdgeBench benchmark without quality loss. The code is distributed under the MIT license as an optional extension for the Pi 0.85.1 harness.

image
image

What happened

The NVIDIA Labs team published the SoL-Pi project, which automates the development of a harness for coding agents: the agents themselves conducted auto-research cycles over 152 proposed hypotheses, of which only four mechanisms survived — roughly one idea out of forty. The search was conducted on 535 environments assembled from 495 GitHub issue-PR pairs and 40 synthetic tasks, and the final validation was performed on the isolated 51-task EdgeBench benchmark, whose data was not used in the search. The surviving mechanisms: Action Fusion combines a file edit and its verification command in a single tool call; Online Context Compact compresses context at the boundaries of completed subtasks; ObservationPack archives large tool output locally, leaving a stable handle in the context with precise page-by-page selection; Evidence-Preserving Reducer delegates reading long logs to a cheaper model, verifying each summary quote against the archive. On EdgeBench with GPT-5.6 Sol and Opus 5 models, the combination maintains the quality of the base Pi, reducing token traffic by 44.7–49.0% and API costs by about a third. The NVlabs/SoL-Pi repository is open under the MIT license with an invitation to contributors to benchmark accepted PRs.

Context

A harness is the environment around the model: tools, context management, and execution loops. In long agent runs, context quickly bloats with tool outputs, and a significant part of the budget goes not to solving the task, but to re-reading already seen data. Auto-research and recursive self-improvement (RSI) are traditionally considered unprofitable: improvement search cycles consume more tokens than they save, so such approaches remained more of a research idea than an engineering practice. SoL-Pi is built on top of the Pi 0.85.1 harness from Earendil Works and does not change the model — all four surviving mechanisms relate to harness systems engineering. The survival rate of about one hypothesis out of forty is informative in itself: the vast majority of "obvious" agent improvement ideas do not pass honest validation. Keeping EdgeBench outside the search data is a standard protection against overfitting improvements to the test set.

Why this matters for the industry

The main result for the industry is not the mechanisms themselves, but a shift in the economics of auto-research: usually such cycles are unprofitable, but here the logic of "before scaling RSI, make AI make AI cheaper" received a working example where improvement selection paid off through reduced run costs. A full improvement selection pipeline for harnesses is demonstrated with strict isolation of held-out validation and gates against metric inflation, and the effect transfers beyond the development environment and is confirmed on two model families, GPT-5.6 Sol and Opus 5, reducing the risk of overfitting to a single model. Practically, this is a ready-made open-source template for saving on long agent runs under MIT, formatted as an optional extension for the Pi harness from Earendil Works — agent frameworks can adopt both individual mechanisms and the selection scheme with benchmarking of accepted PRs. If independent measurements reproduce the numbers, this kind of context engineering — compaction at subtask boundaries, observation archiving with handles, cheap summarization with verification — will start being copied into mainstream harnesses, and "quality per dollar" will stand alongside quality as a metric of competitiveness.

Why this matters for users

If you run coding agents for hours, the effect is directly noticeable: according to the authors, savings reach $8.75–13.50 per hour compared to native Codex and Claude Code, and $4.36–5.71 per hour compared to the base Pi. SoL-Pi is installed on top of unmodified Pi 0.85.1 (requires Node.js 22.19+) with the command pi install git:github.com/NVlabs/SoL-Pi, and mechanisms are enabled one by one in the sol-pi.json configuration. A reasonable implementation order: first local Action Fusion and ObservationPack — they work without additional model calls; then Online Context Compact and Evidence-Preserving Reducer, but only after studying SECURITY.md, since the reducer can send log contents to an external model. The obtained savings should be verified in your own telemetry, not just on the author's benchmark.

What is still unknown / limitations

All numbers so far have been obtained by the authors themselves: there is no independent replication on EdgeBench or analogs. Payback has been demonstrated in one domain — automation of coding agent harness engineering, so it cannot be generalized to "RSI now saves more than it spends"; strictly speaking, this is not RSI, since the model does not improve itself, but improves its harness. EdgeBench is an author-created benchmark, and transferring results to other models, tasks, and load profiles requires verification. Mechanisms with delegation to a cheaper model transfer log contents beyond the local environment, which should be considered when working with confidential code.

Sources

Author

Look at AI, editorial team