Raven: A Harness of Harnesses Built for Recursive Self-Improvement

Raven is introduced as a "harness of harnesses" designed for Recursive Self-Improvement (RSI). As a Host Agent, it integrates built-in and third-party agents to perform complex tasks. Its modular architecture supports the iterative improvement of Raven's own harness by proposing, evaluating, and adopting changes to planning and action mechanisms. The system is powered by EverOS, a local-first memory runtime that maintains context across sessions.

Core Architecture and Agents

Raven employs a modular architecture where the shared harness is refined by the Raven Evolver, a separate tool that evaluates candidate changes against benchmarks. The system includes four built-in agents designed for specific domains, with the text explicitly detailing:

• Raven-Code: Handles agentic software development, including debugging and refactoring.
• Raven-Oncall: Manages unattended workflows and continuous monitoring.

The source states these agents deliver state-of-the-art performance in their respective domains, combining reusable harness components with domain-specific tools and skills.

Self-Evolution Mechanism

The agent loop is decoupled into four strategy modules: Memory, Planning, Capability, and Action. A Curator component rewrites settings or judgment code for these modules. Changes are applied only after passing validation, with automatic rollback if failures occur. The Curator can replace tools, services, skills, and procedures. This experimental feature ships with the repository rather than the installed package.

Ecosystem and Integration

Raven supports orchestration of 13 third-party agents via presets, including Claude Code, Codex, OpenCode, and GitHub Copilot, through ACP, CLI, and OpenAI-compatible APIs. The project is part of the broader EverMind ecosystem, which includes EverOS for memory, SkillForge for skill retrieval from a catalog of 114,190 skills, and various research tools like HyperMem and EverMemBench.