Raven: open-source 'harness of harnesses' turns RSI into working multi-agent engineering, 4.8k stars
aigclink · x · 2026-09-30
The open-source GitHub project Raven (EverMind-AI, 4.8k stars) turns Recursive Self-Improvement from a paper concept into runnable engineering. Raven doesn't do the work itself — it acts as a foreman: decomposing tasks, delegating, and verifying results. It ships with four built-in agents (research, coding, design, unattended runs) and can orchestrate 13 external agents including Claude Code, Codex, and Kimi Code — the author calls it a 'harness of harnesses.'
It self-improves the harness's behavior — planning, tool selection, memory, pre-action checks — in four steps: diagnose failure, propose a change, validate against benchmarks, and keep only changes that beat the original. Crucially, the scoring criteria are immutable to the system itself, closing RSI's biggest trap: an AI grading its own exam.
Related event: Open-Source Raven Turns Recursive Self-Improvement into Working Engineering(3 posts)→
More from coding & agent
- Burkov Reorders ChapterPal's Model Fallback Chain, Crowns GPT-6 Luna Best Value — burkov · 2026-09-30
- Cube Launches Always-On Cloud Computers for Running Claude Code and Codex Agents — algo_diver · 2026-09-30
- A new auto-research loop that bootstraps the shape of the best possible result — burny_tech · 2026-09-30
- Stack Overflow joins OpenAI DevDay to share how Codex sped up its new architecture — pchandrasekar · 2026-09-30
- Models improving doesn't obsolete your agentic coding scaffolding, argues pushback on viral take — max_paperclips · 2026-09-30
- Building a Code Review Agent That Learns From Feedback With Groq and Hindsight — pasulabhavya · 2026-09-30