Google Proposes RRSI: Regularizing Recursive Self-Improvement of Agent Harnesses

Google Research (Han Rujun, Huaxiu Yao, et al., arXiv:2609.24972, submitted 2026-09-21) published "RRSI: Regularized Recursive Self-Improvement of Agent Harnesses," studying how to apply regularization to the recursive self-improvement of agent harnesses (prompts, control flow, tools, memory and context management), with code and project page open-sourced. Conclusion: regularized RSI still yields gains on out-of-distribution tasks, mitigating self-improvement overfitting.

Confirmed

Motivation: LLM agent capability is largely amplified by the harness, and recent methods achieve system-level recursive self-improvement (RSI) by iteratively proposing and selecting component-level edits
The team argues that current "self-improvement" methods—letting agents rewrite their own prompts, tools, memory and workflows—mainly score higher on practiced tasks, with gains often vanishing on new tasks, i.e., overfitting
Elias (omarsar0) relayed the paper's warning: automated harness optimization inflates benchmark scores while real-task performance degrades
RRSI introduces regularization into the recursive self-improvement process to constrain the direction of improvement
Results: 4.7-point gain on OOD tasks and about 30% token savings
Kangwook Lee noted the contrast with weight-space regularization (e.g., L2), which relies on smoothness/simplicity assumptions about functions, whereas RSI calls for harness-level regularization

Why it matters

RSI is reshaping how AI agents evolve, but improvements limited to overfitting trained tasks restrict generalization; RRSI offers a simple, effective regularization scheme and a reference path toward generalizable, self-evolving agent harnesses

2026-09-22 ~ 2026-09-24 · 8 related posts

Primary sources

Google's RRSI regularizes recursive self-improvement of agent harnesses, +4.7 OOD points — google · 2026-09-22 11:17 (UTC+08:00)
Google paper proposes RRSI: regularized recursive self-improvement of agent harnesses — _akhaliq · 2026-09-23 01:37 (UTC+08:00)

[source] Google's RRSI regularizes recursive self-improvement of agent harnesses, +4.7 OOD points — google · 2026-09-22
RRSI: regularized agent self-improvement gains 4.7 points with 30% fewer tokens — HuaxiuYaoML · 2026-09-22
[source] Google paper proposes RRSI: regularized recursive self-improvement of agent harnesses — _akhaliq · 2026-09-23
Paper: RRSI — adding regularization to recursive self-improvement of agents — HuaxiuYaoML · 2026-09-23
Your Self-Improving Agent Is Probably Overfitting, Researchers Warn — HuaxiuYaoML · 2026-09-23
RRSI: Simple Text-Space Regularizers Boost Robustness of Recursive Self-Improving Agent Harnesses — Kangwook_Lee · 2026-09-23
Google paper: auto-optimized agent harnesses overfit evals — RRSI fixes it — omarsar0 · 2026-09-24

1 near-duplicate retellings: cihangxie