Researchers from Google Cloud AI Research and partner universities developed a method called Regularized Recursive Self-Improvement (RRSI) to prevent AI agents from overspecializing on test tasks. The approach focuses on optimizing the agent harness—a framework of prompts, tools, and memory—rather than modifying the underlying language model.
The RRSI method addresses the tendency of self-optimizing agents to memorize benchmarks, which often leads to poor performance on unseen tasks. The system implements a shrinking edit budget to limit changes over time and utilizes a critic to reject proposals that hardcode benchmark-specific solutions or add unnecessary complexity.
In tests across eight benchmarks covering coding, office work, and engineering design, the team used a frozen Claude Opus 4.8 model. RRSI achieved gains of up to 14.1 points on training tasks and up to 4.7 points on unseen benchmarks, while consuming 30 percent fewer tokens at runtime compared to unregularized versions.
The study demonstrated that harnesses optimized with one model can improve others. For instance, a harness refined by Gemini 3.5 Flash increased the accuracy of Gemini 3.1 Flash Lite from 11.2 to 14.6 points. The researchers have made the code for the RRSI method publicly available on GitHub.
