Google Releases RRSI Framework for Self-Improving AI Agent Harnesses

Existing optimization methods for AI agent harnesses suffered from overfitting, where they were excessively tuned to specific benchmarks, causing performance to plummet or even decline in new, out-of-distribution (OOD) environments. The RRSI (Regularized Recursive Self-Improvement) framework released by Google addresses this issue by allowing free modification of harness components while introducing regularization mechanisms into the search loop.

The core regularization mechanisms of RRSI are as follows:

• Edit Budget: Limits the scope of changes to prevent drastic structural modifications.
• Leakage Prevention (Critic): Blocks benchmark logic or ground-truth answers from being directly included in the harness.
• Cost Rule: Rejects changes if the increase in token usage exceeds the allowed range relative to performance gains.
• Noise Floor: Does not adopt minor performance improvements that are not statistically significant.

Experimental results showed that RRSI demonstrated superior transfer performance compared to existing methods across various benchmarks, including Terminal-Bench 2.1 and Harvey LAB. Notably, average performance on unseen new benchmarks (OOD) improved from 39.7% to 43.6%, contrasting with existing methods which saw performance declines in OOD settings. Additionally, on Terminal-Bench, RRSI improved performance from 74.2% to 80.9% by adopting only 5 edits over 20 rounds, demonstrating its efficiency.