Researchers introduced RRSI, a method for improving AI agent harnesses. It adds regularization to recursive self-improvement to prevent overfitting. The proposer limits edit bundles and encourages new paths. The selector uses a critic and pruner to remove weak changes. Tests ran across eight benchmarks in coding, agentic workspaces and engineering design. The method improved scores on both familiar and new tasks. It also reduced the number of policy tokens needed. Code and project page are publicly available.

• up to 14.1 points on in-distribution benchmarks
• up to 4.7 points on out-of-distribution benchmarks
• 30% fewer policy tokens