raven

Raven is an open-source, pre-alpha multi-agent harness from EverMind, described as "the harness of harnesses, built for recursive self-improvement (RSI)." It acts as a host agent that brings built-in and third-party agents together to execute complex tasks, generating DAGs and orchestrating specialized agents within shared workflows. Built-in agents cover four domains: Raven-Research for autonomous deep research and structured reports; Raven-Code for agentic software development, debugging, refactoring and data analysis; Raven-Design for visual deliverables such as slide decks, brand assets, charts and web interfaces; and Raven-Oncall for unattended workflow automation, experimentation and continuous monitoring. The README presents benchmark charts for each, but these are the project's own claims rather than independently verified results. A notable design feature is runtime self-evolution. The agent loop is split into four decoupled strategy modules — Memory, Planning, Capability and Action — which a Curator can rewrite per agent. Changes are verified before installation, and failed checks send them back. The Curator is described as experimental and ships with the repository rather than the installed package. A Persona mechanism assembles a lead role plus specialists from existing agents. Raven connects to third-party agents via ACP, CLI or OpenAI-compatible APIs, with presets for 13 agents including Claude Code, Codex, OpenCode, GitHub Copilot, Qwen Code and others. Core systems include Agent Orchestration, the Evolver, EverOS Memory for cross-session context, SkillForge for skill retrieval, and Proactivity for event monitoring and scheduled execution. Installation options include a curl or PowerShell installer, Docker Compose, or a source checkout. A WebUI is launched with `raven web`, providing chat, task progress, file inspection, and memory/skill browsing. The project is pre-alpha, so interfaces and configuration may change. Documentation, a technical report, and a Chinese README are linked.