RUC open-sources FlashRSI: one framework unifying 13 agent self-improvement algorithms
aigclink · x · 2026-11-02
Researchers from Renmin University of China (RUC NLPIR) open-sourced FlashRSI, a unified framework for agent recursive self-improvement (RSI) that automatically optimizes an agent's prompts, skills, tools, and harness.
- Problem solved: fixing inefficient agent behaviors (re-reading files, skipping tests after bug fixes) normally requires manually iterating on prompts/skills/tool configs — FlashRSI automates propose → edit config → run → evaluate → select → iterate.
- 13 RSI algorithms unified, including OPRO, GEPA, ACE, SkillOpt, Meta-Harness, and RRSI, covering instruction search, skill evolution, context evolution, and harness refinement.
- Reproducible interface: 30 benchmarks × 7 harnesses (MiniAgent, Claude Code, Codex, Pi, Hermes, OpenClaw, OpenCode), with real tasks and benchmarks deciding which changes to keep.
- Full provenance: every round logs which files changed, plus candidate versions, trajectories, scores, and resource usage.
Apache-2.0 licensed on GitHub (RUC-NLPIR/FlashRSI), 200 stars so far.
Related event: RUC Open-Sources FlashRSI, Unifying 13 Agent Self-Improvement Algorithms(3 posts)→
More from coding & agent
- Tinyfish founder on web agents: sites will go agent-friendly, subscriptions don't fit — Saboo_Shubham_ · 2026-11-02
- H3 Continuity Studio: Open-Source ComfyUI Pack Adds Scene Memory to MiniMax H3 — rambrogi · 2026-11-02
- Knowledge Compilation Cuts Agentic Search Query Tokens by 91%, Paper Finds — _reachsumit · 2026-11-02
- Metacognitive Reliability Supervision Teaches LLM Agents to Search Only When Needed — _reachsumit · 2026-11-02
- 10 open-source repos that hack your own app with AI before attackers do — victor_explore · 2026-11-02
- Agent reported "0 results" as success after forgetting a contract — a 5-line memory card fixed it — Humaux · 2026-11-02