RRSI adds search constraints to agent self-modification, with tests showing gains across benchmarks

Google Cloud AI Research and other teams introduced RRSI on September 21, investigating how agents can continually rewrite prompts, tools, and control flow while reducing overfitting to fixed test questions. The method keeps the underlying model weights frozen and applies constraints to the process of generating and selecting modifications. The research code is publicly available. [Paper](https://arxiv.org/abs/2609.24972)

The method first limits how many modifications a single proposal can bundle together, then narrows the scope of changes as evolution progresses, making improvements easier to attribute. The proposer reviews previous hypotheses and results, turning to unexplored components when progress stalls. The selection stage first filters out benchmark-specific logic involving task names, answers, and similar details. It then uses evaluation variability and token costs to constrain candidates, while flagging components that no longer provide benefits for removal. [Method overview](https://regularized-rsi.com/)

The public implementation places each candidate framework in a separate Git worktree and records modification diffs, scores, costs, and adoption decisions. Terminal-based software development, document work, and engineering design each have integration modules and starting frameworks. This lets external teams trace why a particular tool or memory design was retained and rerun the selection process on their own tasks. [Repository](https://github.com/google-research/rrsi)

The study’s main experiments used Claude Opus 4.8. On Terminal-Bench 2.1, which was used for evolution, the coding agent’s score rose from 74.2 to 80.2. Transferring the same framework directly to SWE-bench Verified, which was not involved in selection, increased the score from 82.0 to 83.8. The difference between these gains is a reminder that improvements on the test set used for evolution cannot be treated directly as the extent of generalization. [Public results](https://github.com/google-research/rrsi)

Costs also depend on the baseline. In the ablation experiments on work tasks, RRSI used approximately 2.42 million tokens from the policy model per trial, below the 3.80 million used by unconstrained evolution but still above the original framework’s 1.56 million. Calculations based on Table 2 show a reduction of about 36% relative to unconstrained evolution; the abstract reports approximately 30%. This article uses the tabulated figures and does not extrapolate them into a cost-saving percentage for the entire research and development process. [Ablation table](https://arxiv.org/html/2609.24972v1)

The engineering significance lies in bringing automated modifications into an auditable experimental process and separately examining selection data, held-out tests, and inference costs. The results currently come from the authors’ evaluations, and some work tasks also rely on model-based scoring. Fixed task sets, search budgets, and threshold settings can still affect the outcomes. Further validation should test reproducibility across tool environments and account for the costs of proposal generation, review, and repeated evaluation.