EverMind open-sources Raven, a multi-agent system that evolves a custom harness per model
dair_ai · x · 2026-10-05
EverMind AI released Raven, an open-source multi-agent system that builds a separate harness for each model and domain, evolves those harnesses from their failures, and coordinates everything through a host agent.
- How it works: the host agent splits a goal into a task graph and routes each part to a model-harness pair for research, code, design, or on-call work. Harness components are diagnosed, mutated, and recombined; a candidate replaces the current harness only after passing a statistical check.
- Results: with DeepSeek-V4-Flash, the research harness scores 69.3% on BrowseComp and 60.0% on Humanity's Last Exam, versus at most 62.4% and 43.3% for two other harnesses on the same model.
- Skills: a curated skill library lifts SkillsBench Pass@1 from 9.2% (no skills) to 22.6%.
Paper and Chat-with-Paper links in the original post.
More from coding & agent
- Gemini subscriptions may work in third-party agent harnesses starting December 2 — testingcatalog · 2026-11-22
- Agno's AgentOS guide: build a team agent platform from zero to production — pritisinghhhh · 2026-11-22
- Agents shouldn't know they have memory: background 'dreaming' beats save_memory tools — uriwa · 2026-11-22
- Background memory consolidation tied to compaction, with 3.5k-char budget and prompt-injection safeguards — uriwa · 2026-11-22
- Proactive Grok Bot setup: 28 of 34 weekly stops produce zero notifications — marcelmarais · 2026-11-22
- Anthropic engineer: 90% of staff built self-prompting agent systems around Opus — joemeno · 2026-11-22