![Raven banner](https://github.com/user-attachments/assets/ae944083-9c61-4218-8372-2c3c0d483d4b?raw=true)

X arXiv Hugging Face #1 Paper of the Day Discord WeCom

[Website](https://raven.evermind.ai) · [Documentation](https://evermind-ai.github.io/Raven/) · [中文](README.zh-CN.md)

## ❯❯ What's New ### ❯ Multi-Agent Trajectory Observability **Every agent, every turn, every step — on one timeline.** Raven's new trajectory view renders a running multi-agent session as a live waterfall, from the whole run down to a single step. https://github.com/user-attachments/assets/abb0a5e9-1c37-48bd-9be0-1c9267293337 # What is Raven

Raven one surface, all agents workflow

One Surface, All Agents: Raven generates DAGs and orchestrates multiple specialized agents for complex tasks.

Raven is **the harness of harnesses, built for recursive self-improvement (RSI).** As a **Host Agent**, it brings built-in and third-party agents together to carry out complex tasks. Its modular architecture supports iterative improvement of Raven's own harness: proposing changes to how agents plan and act, evaluating those changes, and adopting improvements that pass validation. Powered by [EverOS](https://github.com/EverMind-AI/EverOS), Raven carries memory and context across sessions to support this process. **Built-in Agents: Raven-Research**, **Raven-Code**, **Raven-Design**, and **Raven-Oncall** support research, coding, visual design, and unattended workflow automation.

Multi-Agent Orchestration Benchmark: Node F1, Edge F1, Partial Order Accuracy, and Exact Match Rate

Raven's Performance on the Multi-Agent Orchestration Benchmark

## ❯❯ Showcase These are three complete projects delivered by Raven. In each case, Raven drove a team of specialized agents from the initial brief or objective through execution to a complete set of final deliverables. ### ❯ THRESHOLD: a complete game development project **The brief came from a person; Raven completed the entire project.** Working autonomously for about 4 days, Raven completed 42 rounds of planning, development, and verification to build a playable first-person shooter in Godot 4, centered on an arena boss fight. The full deliverable includes the game, its poster, presentation, and website, all produced by Raven.

Gameplay video

Website ↗

https://github.com/user-attachments/assets/44724e3e-6564-467a-b422-473aa8f474bd

The THRESHOLD website as one long capture: hero, fight, kill cam, attack tells, evolution, making-of and footer

Poster

Presentation · PPTX ↓

THRESHOLD poster: the Warden towers over the player on a molten arena floor

Cover, slides and closing slide of the THRESHOLD deck

### ❯ Raven RSI: recursive self-improvement in practice **AI that improves AI, with the entire project completed by Raven.** Given a task and evaluation criteria it cannot modify, Raven RSI independently plans each round, writes code, runs experiments, and evaluates the results. In nanochat pre-training experiments, it completed 172 training runs across 7 rounds without a single crash, reducing `val_bpb` by 5.8% within the same 20-minute, single-GPU budget. The same process reduced overshoot in a dam-break simulation by three orders of magnitude and completed an FEA limit-load search in 8 rounds of bisection. The complete deliverable includes the experimental results, visualizations, poster, presentation, and project website, all produced by Raven.

CFD dam-break simulation

FEA limit-load search

Website ↗

Dam-break solve: a collapsing water column resolved to fine free-surface structure, with the out-of-bounds water fraction falling from 1e0 to 1.36e-10 over seven rounds

Limit-load search: a cantilever beam under rising load colored by von Mises stress, with the bisection bracket narrowing from 1800-2000 kN down to 3.125 kN over eight rounds

The Raven RSI website as one long capture: research overview, nanochat, dam-break CFD, FEA solver convergence and model cost comparison

Poster

Presentation · PPTX ↓

Raven RSI poster: a spiral stone stair climbing into the light, with ravens circling it

Cover, slides and closing slide of the Raven RSI deck

### ❯ Raven: a complete product launch project **One raven, a whole flock of specialists. Raven completed the entire project.** Its launch kit brings together a browser-based physics mini-game, a 16-slide product overview, posters in English and Chinese, and the README you are reading now, all produced by Raven.

Physics mini-game · Play online ↗

README

https://github.com/user-attachments/assets/dd186b6d-3752-4c59-89c7-eeea5c6fa962

The top of the Raven README as one long capture: banner, introduction, the four built-in agents with their benchmarks, and runtime self-evolution

Poster

Presentation · PPTX ↓

Raven poster: a raven on a standing stone above sea cliffs at sunset, over the line One raven. A whole flock of specialists.

Cover, slides and closing slide of the Raven overview deck

**[More showcases](docs/showcase.md)** ## ❯❯ Built-in Agents Raven's modular architecture is designed for harness self-evolution and subagent creation. Its four built-in agents deliver **state-of-the-art (SOTA) performance in their respective domains**, combining reusable harness components with domain-specific tools, skills, and agent loops. Raven can delegate a focused task to a single agent or orchestrate multiple agents within a shared workflow. The harness they share is refined by the **Raven Evolver**, a separate tool that consumes Raven as a library and evaluates candidate harness changes against benchmarks; it develops the agents rather than running inside them. > All four agents are built in and ready for orchestration out of the box. ### ❯ Raven-Research **Raven-Research** enables **autonomous deep research** for complex questions, literature reviews, and technical analysis. It delivers clear, structured reports with traceable sources, helping users understand unfamiliar domains, compare alternatives, and make informed decisions.

DeepResearch Mixed: Accuracy, Input Tokens, Output Tokens, and Cost

Raven-Research's performance on the DeepResearch Mixed benchmark

### ❯ Raven-Code **Raven-Code** enables **agentic software development**, turning requirements into working, tested code. It supports feature implementation, debugging, refactoring, data processing, and data analysis, helping users build new capabilities, resolve issues, and improve code quality while following their project's conventions.

Coding Benchmarks: SWE-bench Pro, SWE-bench Verified, WorkBuddy-Code Reward, and SWE-Refactor

Raven-Code's performance on coding benchmarks

DataAgentBench (2026-08-24 Live): Raven-Code with Opus-5 achieves 0.8762 Pass@1

Raven-Code tops on DataAgentBench for data analysis (2026-08-24 Live)

### ❯ Raven-Design **Raven-Design** performs **visual design**, turning ideas and content into polished visual deliverables. It creates PowerPoint slide decks, brand assets, charts, diagrams, and web interfaces, refining layout, typography, and visual consistency to help users communicate clearly and bring their ideas to life.

PresentBench: Raven-Design, Claude Code, and public leaderboard scores

Raven-Design tops on PresentBench for slide generation

Visual Design: Raven-Design, Claude Code, and Hermes on ArtifactsBench Dashboard, ArtifactsBench SVG, and GDPVal

Raven-Design's performance on visual design benchmarks

### ❯ Raven-Oncall **Raven-Oncall** enables **unattended workflow automation** for experimentation, optimization, and continuous monitoring. It autonomously manages workflows from start to completion, sustaining progress over hours or overnight, delivering results, and involving users only when human judgment is needed.

AI4AI (Nanochat 50M Pretraining): Bits Per Byte (BPB), Runtime, Tokens, and Cost

Raven-Oncall significantly outperforms Claude Code on both quality and cost for AI4AI tasks

AI4S Internal Benchmark: Success Rate, Average Total Runtime, Average Total Tokens, and Average Cost

Raven-Oncall significantly outperforms Claude Code on both success rate and cost for AI4S tasks

## ❯❯ Runtime Self-Evolution **Raven is built for this from the ground up.** Its agent loop is split into four decoupled strategy modules — **Memory** for what a turn gets to see, **Planning** for how it approaches the work, **Capability** for which tools an iteration exposes, **Action** for what to do next and for judging it before it runs. A **Curator** keeps rewriting those four seats: a setting, or a small piece of judgement code written for that agent. What it changes belongs to that agent alone, and once installed the agent runs on it. **A Curator changes more than the prompt** — the tools and outside services it reaches for, the skills and procedures it follows, and its own judgement at each point can all be replaced. It keeps going round after round — you put it to work, you say what was wrong, it reworks — until you are satisfied, until it has nothing left worth changing, or until the round budget runs out. Two things hold the quality: nothing is installed before it is verified, and a failed check sends it back; and a round's signals normally reach it with the reference answer stripped out, which limits how directly the answer is exposed. The Curator is experimental: it ships with the repository rather than the installed package. **A Persona is the first thing it builds.** Describe the assistant you want and the Curator assembles one: a lead role that talks to you, and specialists drawn from the agents you already have. What you asked for lands in that assistant's **harness** — its division of work, the tools it may reach for, and the checks it must pass before it acts. > Describe what you need once. Raven assembles the assistant, keeps improving it while you work, and afterwards a sentence is enough to put it to work again. ## ❯❯ Connect Third-Party Agents Raven can connect to and orchestrate agents via ACP, CLI, or OpenAI-compatible APIs, with presets for 15 **third-party agents** and nine **Mirasim-routed** presets to simplify setup, task delegation, and coordination across shared workflows. Try these agents in Raven through a unified interface!

Third-party agents in Raven's Agent Hub

The roster also includes **MiniMax Code** over its native ACP server (`mcode acp`), **DSH** (DeepSeek Harness) over the ACP profile it ships (`dsh --profile acp`), and nine **Mirasim-routed** presets such as `Claude Code-Mirasim`, which run the same agents on Mirasim's accounts and model routes. ## ❯❯ News - 📣 **2026-09-30** — Raven's technical report reached [**#1 on Hugging Face Daily Papers**](https://huggingface.co/papers/2609.33439) — daily (2026-09-30), weekly (2026-10-04) and monthly (2026-10-10).

Three Crowns on Hugging Face: Raven's technical report ranked #1 on the daily, weekly and monthly Hugging Face Daily Papers boards

## ❯❯ Quick Start ### 🤖 Install with Your Agent Let your own agent install Raven for you. Copy this prompt into any agent that can read a web page and run shell commands, such as Claude Code or Codex: ```text Read https://evermind-ai.github.io/Raven/quick-start/ and follow it to install Raven, or to update it if it is already installed. ``` ### 📦 Install Linux, macOS, or WSL2: ```bash curl -fsSL https://raven.evermind.ai/install.sh | bash ``` Native Windows PowerShell: ```powershell irm https://raven.evermind.ai/install.ps1 | iex ``` Windows PowerShell 5.1, the version built into Windows, cannot follow that URL's redirect and stops with `(308) Permanent Redirect`. If you see that error, use the direct installer URL instead: ```powershell irm https://raw.githubusercontent.com/EverMind-AI/Raven/refs/heads/main/install.ps1 | iex ``` Or run it in a container, with nothing on the host but Git and Docker: ```bash git clone https://github.com/EverMind-AI/Raven.git cd Raven docker compose -f docker/docker-compose.yml up ``` Then open `http://localhost:18793` in your browser. Prerequisites are [Git](https://git-scm.com/) and [Docker](https://www.docker.com/) with [Docker Compose](https://docs.docker.com/compose/). See [`docker/README.md`](docker/README.md) for what comes up, how the page signs itself in, and how to run it behind a remotely exposed port. Or install from a source checkout, to develop against the code or to run what has not been released yet: ```bash git clone https://github.com/EverMind-AI/Raven.git cd Raven ./install.sh ``` Run as a file, `install.sh` installs that checkout in editable mode: raven and its bundled plugins link back to your tree, and the TUI bundle and the served page are built from it. A piped run installs the published wheel even from inside a clone, so that a one-line install never picks up whatever a working tree happens to contain. Set `RAVEN_LOCAL_SRC=` to force the editable install through a pipe. The agent products ship with raven itself: a wheel carries the `agents/` product tree and copies it out to your raven home on first use, and a source checkout reads the tree in place. Setup asks, for each product, whether it runs on the model it is tuned for, which needs a key of its own, or on this raven's LLM. Nothing is registered: a folder is on the roster because it is there. See [`agents/README.md`](agents/README.md). Learn more about Raven on the documentation site. **[Read the documentation](https://evermind-ai.github.io/Raven/)** ## ❯❯ Core Systems | System | What it adds | | --- | --- | | **Agent Orchestration** | Coordinates agents, manages task dependencies and parallel execution, and turns multi-step collaboration into reusable workflows. | | **Evolver** | Drives harness self-evolution by diagnosing failures, testing candidate improvements, and retaining changes that outperform the baseline in reproducible evaluations. | | **EverOS Memory** | Preserves user context, agent experience, and world knowledge across sessions, recalling relevant memories and reusable skills for future tasks. | | **SkillForge** | Retrieves relevant skills from local libraries, EverOS memory, and [SkillHub's catalog of **114,190 skills**](https://github.com/EverMind-AI/SkillCorpus#public-artifacts), giving agents specialized expertise on demand. | | **Proactivity** | Combines event monitoring and scheduled execution to anticipate user needs, deliver timely reminders, and initiate follow-up work. |
[![](https://img.shields.io/badge/-Back_to_top-gray?style=flat-square)](#readme-top)
## ❯❯ Launch WebUI Raven's WebUI brings sessions, multi-agent collaboration, and workspace management into your browser. Chat with agents, follow task progress, inspect files and outputs, and browse memory and skills in one place. ```bash raven web ``` The command opens the WebUI in your browser and keeps Raven running in the background. Use `raven web --stop` to stop the background service.

Raven WebUI new session page

New session: one composer, with skills, playbooks, knowledge and memory a click away.

Raven WebUI subagents page

Subagents: every connected agent in one roster, built-in and third-party alike.

## ❯❯ EverMind Ecosystem

The EverMind ecosystem: the EverMind mark and its slogan on an orbital field

[EverMind](https://evermind.ai/) connects memory research, production-ready products, and practical integrations into one open-source ecosystem.
Products
EverOS A local-first, Markdown-native long-term memory runtime for agents and users.
Raven A memory-first, self-improving agent harness with proactivity, context control, and reusable skill retrieval.
EverMe (CLI) A CLI and agent plugin suite for cross-device, cross-agent personal memory.
Research & Evaluation
SkillCorpus Curated, retrieval-ready agent skill corpora with retrieval and evaluation tooling.
EverAlgo Stateless extraction, ranking, parsing, and memory operators that power EverOS.
HyperMem Hypergraph-based hierarchical memory for coarse-to-fine long-term conversation retrieval.
MSA Memory Sparse Attention for scalable latent memory and 100M-token contexts.
EverMemBench Evaluation of factual recall, applied reasoning, and personalized generalization in memory systems.
EvoAgentBench Longitudinal evaluation of agent self-evolution, transfer efficiency, error avoidance, and skill use.
Integrations
OpenClaw OpenClaw plugin for automatic recall, capture, and session-memory lifecycle management.
Hermes Agent Hermes plugin for persistent memory across Hermes sessions.
DeepSeek Harness DSH plugin for memory-aware DeepSeek Harness agents.
Dify Self-hosted and cloud tools for explicit memory search and storage in workflows and agents.
Together, these projects form EverMind's research-to-runtime stack: methods and benchmarks become reusable memory infrastructure, products, and agent integrations.
[![](https://img.shields.io/badge/-Back_to_top-gray?style=flat-square)](#readme-top)
## ❯❯ Contributing Issues and pull requests are welcome. Start with the [developer workflow](docs/dev.md), follow [AGENTS.md](AGENTS.md) for repository rules, and use [GitHub Discussions](https://github.com/EverMind-AI/Raven/discussions) for design conversations. ## ❯❯ License [Apache License 2.0](LICENSE) ## ❯❯ Citation If you use Raven in your research, please cite the [technical report](https://arxiv.org/abs/2609.33439): ```bibtex @misc{evermind2026raven, title = {{Raven: The Harness of Harnesses for Composable Agentic Intelligence}}, author = {{EverMind AI}}, year = {2026}, eprint = {2609.33439}, archivePrefix = {arXiv}, primaryClass = {cs.AI}, url = {https://arxiv.org/abs/2609.33439} } ``` ## ❯❯ Contributors Thanks to everyone who has helped make Raven better. Raven contributors