Raven
The Harness of Harnesses • Super Agent: a trusted, persistent, self-evolving multi-agent ecosystem for all-domain collaboration.
What is Raven?
The Harness of Harnesses • Super Agent: a trusted, persistent, self-evolving multi-agent ecosystem for all-domain collaboration.
Overview of Raven
What it is
Raven is a multi-agent ecosystem described as 'the harness of harnesses,' built for recursive self-improvement (RSI). It acts as a Host Agent that orchestrates multiple specialized agents—both built-in and third-party—to carry out complex tasks. Powered by EverOS, Raven carries memory and context across sessions. Its modular architecture supports iterative improvement of its own harness: proposing changes to how agents plan and act, evaluating those changes, and adopting improvements that pass validation. Built-in agents include Raven-Research (autonomous deep research), Raven-Code (agentic software development), Raven-Design (visual design), and Raven-Oncall (unattended workflow automation).
Key features
- Multi-agent trajectory observability: renders a running multi-agent session as a live waterfall timeline, from the whole run down to a single step.
- Modular architecture with four decoupled strategy modules: Memory, Planning, Capability, and Action.
- A Curator that rewrites those modules—changing prompts, tools, skills, and judgement—and verifies changes before installation.
- Persona creation: describes the desired assistant and the Curator assembles a lead role and specialists from existing agents.
- Connects to and orchestrates third-party agents via ACP, CLI, or OpenAI-compatible APIs, with presets for 15 third-party agents and nine Mirasim-routed presets.
- Built-in agents deliver state-of-the-art performance in their domains: research, coding, design, and unattended automation.
- Recursive self-improvement: Raven can improve its own harness through iterative cycles of proposal, evaluation, and adoption.
Use cases
- Complex multi-agent projects: e.g., THRESHOLD—a complete game development project (first-person shooter in Godot 4) autonomously completed in about 4 days, including the game, poster, presentation, and website.
- Recursive self-improvement experiments: e.g., nanochat pre-training (172 training runs across 7 rounds, reducing val_bpb by 5.8%), dam-break simulation (reduced overshoot by three orders of magnitude), FEA limit-load search (8 rounds of bisection).
- Product launch kit: e.g., a browser-based physics mini-game, 16-slide product overview, posters in English and Chinese, and the README itself.
- General use: deep research, software development, visual design, unattended workflow automation, and orchestrating multiple agents for any domain collaboration.
Getting started
See the repository README for installation and quick-start instructions.
License and pricing
Licensed under Apache-2.0. Open source.
Based on the project's public README and repository metadata.