Show HN: Raven – The harness of harnesses, built for RSI

Raven is a host agent framework that generates directed acyclic workflows and orchestrates specialized agents to solve complex tasks end-to-end. Built as a “harness of harnesses” for recursive self‑improvement, it proposes, evaluates, and adopts changes to its own planning and execution harness via a separate Evolver tool, while EverOS carries memory and context across sessions. The modular stack exposes four decoupled strategy modules - memory, planning, capability, and action - and ships four built‑in agents out of the box: Raven‑Research (autonomous deep research and literature synthesis), Raven‑Code (agentic software development and testing), Raven‑Design (visual deliverables and slide generation), and Raven‑OnCall (unattended workflow automation and continuous experiments).

Demonstrations show fully autonomous project delivery: a playable Godot FPS delivered over ~4 days and 42 planning/development rounds, a product launch kit with web and design assets, and recursive self‑improvement experiments that ran 172 training runs across seven rounds to reduce validation bits‑per‑byte by 5.8% on a single GPU. Physical‑simulation examples include a dam‑break overshoot reduction by three orders of magnitude and an FEA limit‑load bisection converging in eight rounds. Benchmark claims include top results on coding, design, and research suites (e.g., DataAgentBench Pass@1 0.8762) and superior cost/quality vs. comparator agents. The system is pre‑alpha and evolving rapidly.