// the find
EverMind-AI/Raven
The Harness of Harnesses • built for RSI: a trusted, persistent, self-evolving multi-agent ecosystem for all-domain collaboration.
Raven is a multi-agent orchestration harness from EverMind AI: it wraps four built-in agents (research, coding, design, unattended ops) plus third-party CLIs like Claude Code and Codex behind one interface, and adds a memory layer (EverOS) and a 'Curator' that can rewrite its own planning/tooling strategies between runs. It's aimed at people who want one host process coordinating several specialized agents on a single complex task, not at someone who just wants a single coding agent.
The agent loop is split into four separately-replaceable strategy modules (Memory, Planning, Capability, Action), which is a more concrete take on 'self-improving agent' than just rewriting a system prompt. Third-party agent interop (ACP, CLI, OpenAI-compatible API) with presets for 13 existing tools means it's trying to be a coordination layer rather than yet another agent reimplementation. There's an actual benchmarks/ directory in the repo (appworld, clawbench, pinchbench, proactivity_eval) rather than just screenshots, and a Docker Compose path to try it without touching the host. Apache 2.0 license and a CI workflow set (ci.yml, release.yml, commits.yml) suggest some real engineering process behind it.
The README states outright that Raven is pre-alpha with interfaces and config 'changing quickly' — not something to build on top of yet. Every performance claim (SOTA on this benchmark, tops that leaderboard) is backed by an embedded image, not a table or script you can rerun from the README; you have to trust the screenshot. The headline feature — the Curator that does recursive self-improvement — is called out as experimental and explicitly ships outside the installed package, so the thing in the name ('RSI') is the least finished part of the project. The surface area is large (orchestration, memory, skill retrieval, proactivity, evolver, four built-in agents, 13 third-party integrations) for a pre-alpha project, which is a lot of moving parts to debug when something breaks.