OpenRSI scales recursive self-improvement to SGLang-Omni, +7.6%–344% throughput via 14 merged PRs
ChengleiSi · x · 2026-11-10
OpenRSI partnered with SGLang-Omni to take recursive self-improvement (RSI) from toy demos to production AI infrastructure:
- How it works: agents autonomously hypothesize, design experiments, and iterate; an independent verifier validates every gain, while humans only define problems and review the final code
- Results: the loop ran across 8 omni models, raising H100 throughput by +7.6% to +344.3% without touching model weights or slowing other models
- Shipping: 14 PRs have passed upstream review and been merged by SGLang-Omni maintainers
Each merged optimization becomes the baseline for the next round. RSI-Harness and the RSI-Anything data pipeline are open-sourced.
More from coding & agent
- Two-model workflow: Fable 5.2 for planning, Opus 5.5 for execution — daniel_mac8 · 2026-11-10
- One hour with Fable 5.2: plan with one model, execute with another — daniel_mac8 · 2026-11-10
- Replit launches Monitoring Agent to auto-detect, fix and deploy fixes for vibe-coded apps — amasad · 2026-11-10
- Multi-agent harnesses keep beating single agents on accuracy and latency in ProgramBench — marktenenholtz · 2026-11-10
- fx: a tiny open-source coding agent in Zig with a 5.97 MiB single binary — evilrabbit_ · 2026-11-10
- 20 Python MCP Tools Worth Knowing, From Server Building to Gateways, Security and Eval — MaryamMiradi · 2026-11-10