EverMind-AI/Raven

The Harness of Harnesses • built for RSI: a trusted, persistent, self-evolving multi-agent ecosystem for all-domain collaboration.

What it solves

Raven acts as a "host agent" that orchestrates multiple specialized AI agents to complete complex, multi-step projects. It solves the problem of fragmented AI tools by providing a single interface to manage diverse agents—both built-in and third-party—and allows these agents to recursively improve their own planning and execution strategies over time.

How it works

Raven generates Directed Acyclic Graphs (DAGs) to plan and coordinate workflows across different agents. It utilizes a modular architecture divided into four strategy modules: Memory, Planning, Capability, and Action. A separate tool called the Curator iteratively rewrites these modules (including tools, skills, and judgment code) based on performance and user feedback, ensuring that improvements are verified before being adopted. It is powered by EverOS to maintain memory and context across different sessions.

Who it’s for

It is designed for users needing autonomous execution of high-complexity tasks, such as full-scale game development, deep technical research, software engineering, and unattended workflow automation (AI4AI/AI4S).

Highlights

  • Built-in Specialized Agents: Includes Raven-Research (deep research), Raven-Code (software development), Raven-Design (visual assets/slides), and Raven-Oncall (unattended automation).
  • Recursive Self-Improvement (RSI): Capable of independently planning, coding, and evaluating its own performance to optimize results without human intervention.
  • Third-Party Integration: Supports orchestration of external agents via ACP, CLI, or OpenAI-compatible APIs, with presets for 13 different agents.
  • Runtime Self-Evolution: Uses a Curator to modify agent personas and harness configurations dynamically based on task outcomes.

Related

  • Dispatch
    Raven: The Harness of Harnesses for Recursive Self‑Improvement

    Raven is an open‑source, memory‑first multi‑agent platform that orchestrates built‑in and third‑party agents, supports runtime self‑evolution, and demonstrates end‑to‑end project delivery and recursive self‑improvement benchmarks.

  • Project
    raven-ml/raven

    Raven is an alpha‑stage OCaml ecosystem that re‑creates the Python data‑science/ML stack (NumPy, JAX, PyTorch, pandas, Matplotlib, etc.) with type‑safe libraries for arrays, autodiff, neural nets, optimizers, tokenisation, dataframes, plotting, notebooks and experiment tracking.

  • Project
    Laxcorp-Research/project-raven

    An open-source meeting copilot that provides real-time AI assistance and automated recaps using local storage and user-provided API keys for transcription and LLMs.

  • Project
    nageoffer/ragent

    Ragent AI is an open‑source, production‑ready Java platform that implements a full RAG pipeline and an Agentic ReAct engine. It offers hybrid retrieval (vector, keyword, graph, web), advanced question understanding, model tiering with fault tolerance, session memory, Redis‑based traffic protection, and a complete React admin console. The code is split into seven Maven modules (infra‑ai, system, rag, agent, bootstrap, framework, mcp‑server) to keep AI provider logic separate from business logic, making it easy to swap models, vector stores, or tools. Designed for Java backend engineers, it provides real‑world engineering patterns (circuit‑breaker, transactional MQ, distributed rate limiting, observability) and serves as a solid reference for building enterprise AI services.

  • Project
    Emiyaaaaa/HiveMind

    Hivemind is a production‑ready runtime for multi‑agent systems. It adds durable run history, ordered step/message logs, SSE/webhook event streaming, and a pluggable adapter layer (defaulting to LangGraph) on top of existing agent frameworks. The stack combines a Java Spring Boot API, Python asyncio workers, PostgreSQL persistence, Redis job/event bus, and a Next.js admin console, all under an Apache‑2.0 license.