# google-research/rrsi

> Research framework for automatically evolving an LLM agent's harness (prompts, tools, memory, control flow) with regularized search to avoid overfitting to the training tasks.

- Magnitude: 4.4 out of 10 — Breakout
- Stars: 1,016 total · +61 stars today, ≈ 110 by evening
- Star trust: star growth looks organic
- Category: AI agents · Language: Python · License: Apache-2.0 · Created: 2026-09-16 · Last push: 2026-09-23
- GitHub: https://github.com/google-research/rrsi · Page: https://gitnova.dev/en/r/google-research/rrsi

## Useful for

- Run harness evolution on your own task set via python3 rrsi.py --domain run
- Evaluate the base harness and seed the frontier using the baseline mode
- Check the evolved harness's robustness on out-of-distribution tasks

## Why it’s here

- 61 stars so far today, about 110 expected by the end of the day. The usual pace is 15 per day, so that's 7.3× as much.
- Over the last two days the pace is 4.7× that of the previous week and a half.
- The spike has held for 2 days in a row — not a one-off blip.
- The repository is 14 days old and already has 1,016 stars.
- Top new repositories this month: #98.
- Hacker News: “RRSI: Regularized Recursive Self-Improvement of Agent Harnesses” — 9 points, 1 day ago.
- About 15 forks a day — people are taking the code.

## Star trust

Star growth looks organic.

Star-trust labels are heuristics based on the repository’s behavior, not a check of every stargazer.

## Numbers

- Forks: 84
- Issues and pull requests: 4
- Watchers: 3
- Average over the last week: 114 per day
- Usual pace: 15 per day
- Stars in the last hour (measured): 7

## Stars per day, last 18 days (oldest → newest, today is partial)

2026-09-13 … 2026-09-30: 0, 0, 0, 0, 0, 0, 0, 0, 22, 129, 119, 53, 22, 13, 35, 343, 223, 61

## Hacker News

- RRSI: Regularized Recursive Self-Improvement of Agent Harnesses — 9 points, 0 comments: https://news.ycombinator.com/item?id=49881797

## Spotted in now

- Top new repositories this month: #98

## More in this category

1. **NVIDIA/OpenShell** — 8.4 · Breakout · AI agents · Rust · +825 stars today, ≈ 1,265 by evening
   A safe, private runtime for autonomous AI agents: isolated sandbox containers governed by declarative YAML policies restricting filesystem, network, and process access. Manages credential injection for agents like Claude, Codex, OpenCode,…
   Full card: https://gitnova.dev/en/r/NVIDIA/OpenShell.md
2. **feder-cr/dots** — 8.3 · Early signal · AI agents · Python · +1,102 stars today, ≈ 1,546 by evening
   Open-source web AI agent with its own patched Firefox browser that avoids detection and blocking. The model is swappable via OpenRouter, while the browser keeps a consistent identity and persistent sessions.
   Full card: https://gitnova.dev/en/r/feder-cr/dots.md
3. **rehan-remade/universal-modder** — 7.3 · Early signal · AI agents · Python · +413 stars today, ≈ 558 by evening
   A set of skills, tools and the fal MCP for Claude Code that lets it mod almost any PC game: engine recon, reverse engineering, asset generation, in-game testing and showcase video recording.
   Full card: https://gitnova.dev/en/r/rehan-remade/universal-modder.md
4. **paperclipai/paperclip** — 7.3 · Peaking · AI agents · TypeScript · +350 stars today, ≈ 643 by evening
   Open-source Node.js and React platform for orchestrating a team of AI agents: goals, org chart, budgets, tickets and audit. Suits teams running autonomous agents (Claude Code, Codex, Cursor, etc.) and wanting to manage them like employees.
   Full card: https://gitnova.dev/en/r/paperclipai/paperclip.md
5. **yetone/magpie** — 7.2 · Breakout · AI agents · Go · +461 stars today, ≈ 802 by evening
   A menu-bar utility for macOS/Linux/Windows that lets you pick the model for each AI agent (Codex, Claude Code, Gemini CLI, etc.) and runs a local gateway translating OpenAI and Anthropic APIs between agents and providers.
   Full card: https://gitnova.dev/en/r/yetone/magpie.md

---

Magnitude (0–10) measures how fast and how unusually interest in a repository is growing right now. It is not a quality score.
Days are UTC. “So far today” is a fact; “expected by the end of the day” is a forecast.
Summaries and use cases are written by an LLM (DeepSeek V4.1 Flash) from the README and may be inaccurate: verify specific claims (benchmarks, speed, hardware) in the repository itself.
Data as of 2026-09-30 15:35 UTC, updated every 30 minutes.