RUC open-sources FlashRSI: unified framework for 13 agent self-improvement algorithms
aigclink · x · 2026-11-02
RUC's NLPIR team open-sourced FlashRSI, a unified framework for RSI (Recursive Self-Improvement) research that lets agents automatically optimize their own prompts, skills, tools, and harness.
Example: a coding agent that repeatedly re-reads files wasting tokens, or forgets to run tests after fixing bugs, normally requires manual prompt/config tweaking. FlashRSI automates the loop: propose improvement → modify agent config → run tasks → evaluate → select better version → iterate.
Key features:
- Unifies 13 self-improvement algorithms including OPRO, GEPA, ACE, SkillOpt, Meta-Harness, and RRSI, covering prompt, skill, experience memory, and harness optimization layers
- Supports 30 benchmarks and 7 harnesses (MiniAgent, Claude Code, Codex, Pi, Hermes, OpenClaw, OpenCode)
- Each run records which prompts/skills/tools/configs changed, keeps candidate versions and snapshots, and uses real tasks/benchmarks to decide what to keep
Apache-2.0 licensed, 200 GitHub stars, with English and Chinese docs.
Related event: RUC Open-Sources FlashRSI, Unifying 13 Agent Self-Improvement Algorithms(3 posts)→
More from coding & agent
- Agent drives Final Cut Pro to edit a full music video end-to-end — ZeroStateReflex · 2026-11-02
- Zed CEO shows off SLAG, a 1-8 player party platformer built with Codex — zeeg · 2026-11-02
- Grok Bot tip: have it write a Python script and run it on a routine to save usage — nima_owji · 2026-11-02
- mnemos v2 MCP nearly ready: watch your agent's memories form live in an in-app browser — RileyRalmuto · 2026-11-02
- Supabase launches Compute in private alpha: run agents next to Postgres — dshukertjr · 2026-11-02
- ChatGPT Sites packs login, built-in database, analytics and MCP hosting — gabrielchua · 2026-11-02