SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness

Problem

This work addresses a gap in token efficiency for scaling recursive self-improvement in coding agents. The authors propose a novel approach to enhance the performance of auto-research loops, which is particularly relevant given the increasing demand for efficient coding agents in AI applications. The paper is a preprint and has not yet undergone peer review.

Method

The proposed method, SoL-Pi, is inspired by recursive self-improvement (RSI) principles and focuses on scaling auto-research loops. Key mechanisms include:

• Action Execution: Streamlining the process of executing actions within the agent's environment.
• Context Compaction: Reducing the amount of contextual information required for decision-making, thereby improving efficiency.
• Observation Handling: Enhancing the way observations are processed to facilitate better learning and adaptation.
• Delegated Reading: Allowing the agent to delegate certain reading tasks to optimize information retrieval and processing.

Results

The evaluation of SoL-Pi was conducted using the EdgeBench benchmark, where it demonstrated comparable performance to the Pi model across GPT-5.6 Sol and Opus 5. Notably, the method achieved a token traffic reduction of 44.7-49.0% compared to native Codex and Claude Code harnesses. Additionally, it resulted in an API cost reduction of approximately one third relative to these baselines. The estimated hourly savings were reported to be between $8.75 and $13.50 when compared to native Codex and Claude Code harnesses, and between $4.36 and $5.71 when compared to Pi.

Limitations

The authors did not report any limitations in the study. However, as with any preprint, the lack of peer review may mean that certain aspects of the methodology or results could be subject to scrutiny or revision.

Why it matters

The implications of this work are significant for the development of more efficient coding agents capable of recursive self-improvement. By enhancing token efficiency and reducing operational costs, SoL-Pi could facilitate broader adoption of AI coding agents in various applications, potentially leading to advancements in automated programming and software development. This research opens avenues for further exploration into optimizing agent architectures and improving their learning capabilities.