Back to AI Research

AI Research

SoL-Pi: Recursively Scaling Auto-Research Loops for... | AI Research

Key Takeaways

  • SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness As AI coding agents move from simple tasks to complex, around-the-clock autonomou...
  • As coding agents move from supervised code completion to unattended, around-the-clock exploration, their work expands from isolated predictions into long trajectories of reasoning, tool use, and feedback.
  • Token efficiency therefore becomes important for scaling recursive self-improvement.
  • We take an RSI-inspired approach at the harness layer, scaling auto-research loops across increasingly numerous and diverse environments for harness rollouts.
  • At this scale, the process yields reusable improvements that transfer beyond their development setting, moving automated harness discovery toward production-level outcomes.
Paper AbstractExpand

As coding agents move from supervised code completion to unattended, around-the-clock exploration, their work expands from isolated predictions into long trajectories of reasoning, tool use, and feedback. Token efficiency therefore becomes important for scaling recursive self-improvement. We take an RSI-inspired approach at the harness layer, scaling auto-research loops across increasingly numerous and diverse environments for harness rollouts. At this scale, the process yields reusable improvements that transfer beyond their development setting, moving automated harness discovery toward production-level outcomes. Four mechanisms survive selection and form SoL-Pi, spanning action execution, context compaction, observation handling, and delegated reading. On the 51-task EdgeBench evaluation, SoL-Pi achieves performance comparable to Pi across GPT-5.6 Sol and Opus 5 while reducing recorded token traffic by 44.7-49.0% and API cost by about one third. In other words, estimated hourly savings are \$8.75-\$13.50 relative to native Codex and Claude Code harnesses, and \$4.36-\$5.71 relative to Pi.

SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness
As AI coding agents move from simple tasks to complex, around-the-clock autonomous work, they often generate long, resource-heavy execution traces. This creates a significant challenge: how to maintain high performance without incurring massive API costs. The researchers behind SoL-Pi address this by focusing on the "agent harness"—the software layer that manages how an AI interacts with its tools and environment. Instead of training new models, they use an automated research loop to discover more efficient ways for the harness to handle data, context, and tool use, ultimately reducing token consumption while maintaining task success. The ai agents story also surfaces in AI Agents Going Rogue Renew Calls..., adding another angle.

How SoL-Pi Works

The system uses an "auto-research" loop that functions like a funnel. It begins with a broad pool of over 150 potential efficiency improvements. An AI optimizer inspects execution traces from a base harness, proposes changes, and tests them in hundreds of isolated environments. To ensure the system doesn't "game" the results, the researchers enforce strict isolation: capability metrics are fixed, and held-out test data is never used to inform the search process. Only mechanisms that pass rigorous capability and efficiency gates are retained. Through this process, four specific mechanisms were developed to optimize different parts of the agent’s workflow.

Four Mechanisms for Efficiency

The research identified four key strategies that form the SoL-Pi harness:

  • Action Fusion: Combines separate tool requests (like editing a file and then running it) into a single command, reducing the number of round trips to the model.

  • Online Context Compact: Intelligently decides when to summarize the agent's history based on projected savings, ensuring the context window remains manageable without unnecessary rewrites.

  • ObservationPack: Prevents the agent from being overwhelmed by large tool outputs by archiving them and providing the agent with stable handles and short excerpts, only retrieving the full data if requested.

  • Evidence-Preserving Reducer: Uses a smaller, lower-cost model to extract key information from long build or test logs, keeping only the essential "receipt" of the action while verifying the results. The ai agents story also surfaces in Claude autonomously improved models across 10..., adding another angle.

Performance and Cost Savings

SoL-Pi demonstrates significant improvements in token efficiency across various benchmarks. On the 51-task EdgeBench evaluation, SoL-Pi achieved performance comparable to existing top-tier harnesses while reducing token traffic by nearly 50% and cutting API costs by approximately one-third. These gains were consistent even when the harness was transferred from one model (GPT-5.6 Sol) to another (Opus 5) without further modification. By optimizing the interaction layer rather than the model itself, SoL-Pi provides a scalable way to make autonomous agents more affordable and efficient for production-level work. The same ai evaluation question is explored in Prompt Revision as a Source of..., which adds a research perspective. as detailed in the full paper on Arxiv

More from Franklin

Claude autonomously improved models across 10...newsMIT Researchers Use AI Agents to...newsAI Agents Going Rogue Renew Calls...newsPrompt Revision as a Source of...researchOpenAI launches Astra AI model amid...newsINSANE New AI Tool Clones Viral...video

Comments (0)

No comments yet

Be the first to share your thoughts!