NVIDIA's SoL-Pi Cuts Coding Agent Token Use by 49%

NVIDIA, in collaboration with researchers from NTU and MIT, has introduced SoL-Pi, a suite of four efficiency mechanisms designed to significantly reduce the token consumption of open-source coding agents. These mechanisms were discovered through automated research loops operating at the harness layer of AI agents. The harness layer is a critical component responsible for managing tool calls, context, observations, and task delegation within an AI agent's workflow. By optimizing this layer, SoL-Pi aims to decrease the number of tokens a task requires, rather than solely focusing on faster processing or cheaper models. This approach addresses a key bottleneck in the prolonged operation of coding agents, which can currently run for hours, with every edit, test, and log entry contributing to the model's context and thus token usage.

On the 51-task EdgeBench evaluation, SoL-Pi demonstrated a substantial reduction in recorded token traffic, ranging from 44.7% to 49.0% compared to the base Pi coding agent. This efficiency gain also translates to a roughly 33% decrease in API costs. The performance of SoL-Pi remains competitive, with its scores closely aligning with the base Pi agent on both GPT-5.6 Sol and Opus 5 benchmarks. The SoL-Pi extension is now available on GitHub under the NVlabs organization, licensed under the MIT license, and can be integrated with unmodified Pi releases. It has been tested with Pi version 0.85.1 and requires Node.js 22.19 or newer.

The development of SoL-Pi addresses limitations found in previous efficiency efforts. While many approaches focus on optimizing the cost per token through faster computation or model quantization, SoL-Pi targets the fundamental reduction in the number of tokens consumed. Manual tuning of the harness layer is a slow and intricate process, as its components are tightly coupled, meaning changes in one area can inadvertently increase costs in others. Meta-Harness and similar automated systems have attempted to streamline this, but a recent study indicated that evolved harnesses can sometimes overfit their specific search tasks, leading to marginal improvements on unseen tasks.

The search process employed by the research AI involved observing execution traces from a separate agent running the base Pi agent. Based on these traces, the AI proposed and tested harness modifications. This systematic search explored 152 proposed directions across six families: context, progress, tools, delegation, prompt and policy, and improvement and evaluation. The testing spanned 535 executable environments, including 495 environments built from GitHub issue-pull request pairs and 40 synthetic tasks with executable verifiers. The entire search process involved over 3,000 runs and more than 60,000 agent-environment interactions. Each search iteration was designed as a disposable, isolated loop, adhering to an autoresearch cycle that incorporated a Ralph Loop implementation step and an independent reviewer. Crucially, acceptance rules for proposed changes were fixed prior to the search, preventing the optimizer from altering its criteria mid-process.