Raven 0.2.0 pitches a self-rewriting orchestration layer
Raven 0.2.0 argues its harness layer can be rewritten, with prompts, policies, playbooks and the orchestrator all iterable across multi-day task loops.
A model gets smarter when you train it. A team gets smarter when you can rewrite the way it works. That’s the 0.2.0 bet. Raven is not trying to be the best researcher, coder, designer, and on-call lead in one window. It hosts specialists its own Research, Code, Design, Oncall, plus Claude Code and Codex on one task graph with shared memory. 0.963 Node F1 on multi-agent orchestration. Days-long loops, not one-shot chats. The RSI claim sits above the weights. Prompts, policies, playbooks, strategy code, even the orchestrator, are instances that can be improved. Same base model on the benches so the harness is what’s being measured. Then 172 nanochat runs, 7 rounds, no crash, val_bpb down 5.8% on a locked 20-minute single-GPU budget. Open source, Apache-2.0. A lot of the launch material was made in Raven. The scaffolding shipped the announcement, then offered to rewrite itself.
This post quotes another post: x.com/longtermmemorye/status/2104745468119146913