Google Research open-sources RRSI for self-improving AI agents without overfitting
Google Research, alongside UNC-Chapel Hill, Stanford, and Washington University in St. Louis, has released RRSI (Regularized Recursive Self-Improvement), an open-source framework that lets AI agents refine their own harness—prompts, tools, memory, and workflows—without altering model weights. The method prevents overfitting by constraining the improvement loop, ensuring gains transfer beyond the…
Key points
- RRSI lets AI agents edit prompts, tools, and workflows without changing model weights, preventing overfitting
- Benchmarks show Terminal-Bench 2.1 improved from 74.2% to 80.2% and SWE-bench Verified from 82.0% to 83.8%
- Open-source framework uses Apache 2.0 license, requires Python 3.10+, and defaults to Claude Opus 4.8 on Vertex AI
RRSI introduces mechanisms like an annealed edit budget, evidence-aware credit, and a leakage critic to filter out memorization and noise. Benchmark results show improvements across eight datasets, including Terminal-Bench 2.1 (74.2% → 80.2%) and SWE-bench Verified (82.0% → 83.8%), with out-of-distribution gains on JobBench (+4.7), GDPval (+3.5), and APEX-Agents (+3.7). The framework reduces policy token usage by 30–36% compared to unregularized evolution. The code is Apache 2.0-licensed and works with any LiteLLM-compatible model, defaulting to Claude Opus 4.8 on Vertex AI.
The story so far
6 episodes →- Google Research open-sources RRSI for self-improving AI agents without overfittingthis story
Google Research Open-Sources RRSI: AI Agents That Improve Their Own Harness Without Overfitting
MarkTechPost · 29 September 2026
Loading the full article…
This text was published by MarkTechPost and written by Asif Razzaq. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
3sources- Self-Evolving Harness on Multiple Tasks with the Agent as Its Own OptimizerPrimary source · arXiv cs.AI ·
- SAGE: A Statistical Acceptance Gate for Self-Evolving AgentsPrimary source · arXiv cs.AI ·
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Agents & Tools
All →- Gemini may delete existing text when asked to append, author reports · 1 src
- Google's Gemini outperforms ChatGPT and Claude as Android assistant · 1 src
- Meta launches Muse AI agent in Mexico, first in Latin America · 11 src
- Meta rolls out real-time voice and expressive avatars for Muse AI in the US · 7 src
- Meta improves Muse and Meta AI with Active Memory and Active Focus · 2 src
Comments
via GitHub Discussions