⬢github Python · 880 ★ +214 since we first saw it · pushed 5 d ago · Apache-2.0
google-research/rrsi
RRSI is a Google Research Python framework that automatically improves LLM agent harnesses—the prompts, tools, control flow, and memory around a frozen model—via a regularized search loop. It generates candidate harness edits in git worktrees, screens them with a critic, evaluates them against benchmarks, and accepts gains only when they exceed noise, avoiding overfitting to the evolution task set.
Why now: The paper and project page just came out (September 2026) and the repo is trending on GitHub's most-starred new repos, drawing attention to the recursive self-improvement approach.
Who it is for: ML researchers and engineers building LLM agents who want to automatically optimize and evolve agent scaffolding beyond hand-tuned harnesses.
Stars over our 55 snapshots: 666 to 880, since 13 h ago.
Where people talked about it
- ⬢github new repos, most starred just now
API: https://socialmediatrends-api.osmike.com/v1/repos/google-research/rrsi