# Google Research releases RRSI framework for self‑improving AI agents

> The open‑source tool lets large‑language‑model agents adjust prompts, tools and memory without changing the model itself, boosting benchmark scores.

Oossa · 2026-09-29 · https://oossa.com/en/google-research-releases-rrsi-framework-for-self-improving-ai-agents

Google Cloud AI Research has made RRSI publicly available. The framework lets LLM agents rewrite their own prompts, tools and memory while keeping the model weights frozen. It adds safety checks like a leakage critic and a cost rule to keep improvements from overfitting. In tests with Claude Opus 4.8, the Terminal‑Bench 2.1 score rose from 74.2% to 80.2%, and all six held‑out splits showed gains.

## The facts

- RRSI released Sep 29, 2026
- Terminal‑Bench 2.1 score improved to 80.2% with Claude Opus 4.8

## Why it matters

It shows how AI agents can get better at new tasks without retraining the underlying model, saving compute and reducing overfitting risks.

## Sources & references

1. [Google Research Open-Sources RRSI: AI Agents That Improve Their Own Harness Without Overfitting](https://www.marktechpost.com/2026/09/29/google-research-open-sources-rrsi-ai-agents-that-improve-their-own-harness-without-overfitting/) – MarkTechPost, 2026-09-29

Last updated: 2026-09-29