引言
"The Harness of Harnesses."
这是"一天一个开源项目"系列的第 245 篇文章。今天带你了解的项目是 Raven（EverMind-AI/Raven）。
最近一年，Claude Code、Codex、OpenCode、Hermes Agent 这类 coding / 通用智能体层出不穷，每个都有自己的 harness（循环、工具、记忆、技能的组合方式）。问题随之而来：这么多智能体，谁来调度它们？谁来决定哪个任务交给哪个？harness 本身又能不能自己变好？
Raven 给出的定位很直白：它是一个 Host Agent，把内置智能体和第三方智能体组合起来完成复杂任务，并且通过模块化架构，让 harness 本身也成为可以被迭代改进的对象。
你将学到什么
- Raven 的定位：为什么叫 'Harness of Harnesses'
- 四个内置智能体（Research / Code / Design / Oncall）各自负责什么
- 运行时自进化：Agent loop 拆成哪四个解耦策略模块，Curator 做什么
- 五个核心系统：编排、Evolver、EverOS 记忆、SkillForge、Proactivity
- 如何安装并启动 WebUI
- 它适合谁、局限在哪
前置知识
- 了解 LLM Agent 的基本循环（规划 → 工具调用 → 观察）
- 用过 Claude Code / Codex 之类的编码智能体
- 对 MCP / ACP 等智能体接入协议有基本印象更佳
项目背景
项目简介
Raven 自称"面向递归自我改进（RSI）的多智能体生态"。它做三件事：
- 宿主与编排：作为 Host Agent，协调多个智能体、依赖和并行。
- 自我改进：模块化架构允许迭代改进自身 harness——提出改动、评估、采纳通过验证的改进。
- 持久记忆：记忆层由 EverOS 提供，上下文可以跨会话保留。
作者/团队介绍
- 团队：EverMind-AI（EverMind AI）
- 技术报告：Raven: The Harness of Harnesses for Composable Agentic Intelligence，arXiv 2609.33439（项目页称该论文获 Hugging Face Paper of the Day 第一，2026-09-30）
- 生态：EverMind 同时有 EverOS（记忆）、EverMe（CLI），以及 SkillCorpus、EverAlgo、HyperMem、MSA、EverMemBench、EvoAgentBench 等研究与评测项目，并与 OpenClaw、Hermes Agent、DeepSeek Harness、Dify 插件有集成
- 社区：X、Discord、WeCom
- 项目创建时间：项目页未明确给出
项目数据
- ⭐ GitHub Stars: 5.7k
- 🍴 Forks: 198
- 📝 Commits: 1,910
- 📄 License: Apache-2.0
- 🌐 官网: raven.evermind.ai
- 📚 文档: evermind-ai.github.io/Raven
主要功能
核心作用
Raven 想解决的是"多智能体怎么组合、harness 怎么变好"这两个问题：一方面把不同智能体当作可组合的能力单元，由 Host 统一调度；另一方面把 Agent loop 里的关键策略拆开，让它们可以被替换、评测和进化，而不是写死在某个产品里。
使用场景
- 自主深度研究
- Raven-Research 输出带可追溯来源的结构化报告。
- 智能体式软件开发
- Raven-Code 负责功能实现、调试、重构，以及数据处理与分析。
- 视觉设计产出
- Raven-Design 生成 PPT、品牌素材、图表、网页界面。
- 无人值守的长时任务
- Raven-Oncall 做实验、优化、持续监控，可以持续数小时甚至通宵。
- 跨智能体协作
- 把 Claude Code、Codex 等已有智能体作为第三方 worker 接入，由 Raven 编排。
快速开始
最省事的方式是让你已有的智能体帮你装，把下面这句话交给它：
Read https://evermind-ai.github.io/Raven/quick-start/ and follow it to install Raven, or to update it if it is already installed.
手动安装：
# Linux / macOS / WSL2
curl -fsSL https://raven.evermind.ai/install.sh | bash
# Windows PowerShell
irm https://raven.evermind.ai/install.ps1 | iex
Docker 方式（需要 Git 和 Docker Compose）：
git clone https://github.com/EverMind-AI/Raven.git
cd Raven
docker compose -f docker/docker-compose.yml up
# 然后访问 http://localhost:18793
启动 WebUI：
raven web # 后台运行并打开浏览器
raven web --stop # 停止
WebUI 支持对话、多智能体协作、工作区管理、查看文件与输出、浏览记忆和技能。
安装脚本直接运行为可编辑安装；通过管道执行则安装已发布的 wheel。安装时会让你为每个智能体产品选择：使用其调优模型（需要独立密钥），还是使用 Raven 自身配置的 LLM。
核心特性
- 四个内置智能体
- Research、Code、Design、Oncall，覆盖研究、开发、设计、长时自动化。
- 13 个预设第三方智能体
- 通过 ACP、CLI 或 OpenAI 兼容 API 接入：Claude Code、Codex、OpenCode、Hermes Agent、OpenClaw、MiroThinker、GitHub Copilot、Qwen Code、CodeBuddy、Qoder、Grok Build、Kimi Code、Pi。
- Evolver：评测驱动的 harness 进化
- 诊断失败、测试候选改进，只保留优于基线的改动。
- 运行时自进化（Curator）
- 持续改写 Agent loop 的四个策略模块。
- EverOS 跨会话记忆
- 保存用户上下文、经验与知识。
- SkillForge
- 从本地库、EverOS 记忆和 SkillHub 检索技能（项目页给出的 SkillHub 规模为 114,190 个技能）。
- Proactivity
- 事件监控与定时执行，主动提醒和跟进。
- Persona
- 描述想要的助手，Curator 组装一个主角色加若干专家。
项目优势
项目页没有给出与同类项目的正式对比，下面是我基于其定位做的定性比较，仅供参考：
为什么值得关注？
- 它把"harness 是可以被工程化改进的对象"这件事做成了产品形态，而不是停留在论文里。
- 它不要求你放弃手里的 Claude Code / Codex，而是把它们纳入编排。
- 改进有门槛：项目页强调验证前不安装、失败则返工，并且信号通常会剥离参考答案。
项目详细剖析
整体结构：Host + Workers + 五个系统
Raven 的核心系统可以归纳为五块：
内置智能体
这些数字来自项目自述，没有经过我独立复现，读者引用前建议回到论文和 benchmarks 目录核对。
共享的 harness 由 Raven Evolver 改进。Evolver 是独立工具，以库的方式使用 Raven，用基准评估候选改动。
Agent loop 的四个解耦模块
这是 Raven 设计里最值得细看的一点。它把 Agent loop 拆成四个策略模块：
- Memory：记什么、怎么取
- Planning：怎么拆解任务
- Capability：有哪些工具、技能
- Action：怎么执行与判断
四块解耦之后，每一块都可以被单独替换和评测。Curator（实验性，随仓库发布，不在安装包内）会持续改写这四处，可以替换工具、技能、流程和判断逻辑。
这和我们在评测系列里反复强调的思路一致：想让一个系统可持续变好，首先要让它的可变部分有清晰的边界，其次要有可信的评测来决定改动要不要留下。Raven 的质量保障规则（验证前不安装、失败返工、信号剥离参考答案）本质上就是在给自我改进加评测门控。
展示案例
项目页给了三个案例：
- THRESHOLD：一个 Godot 4 第一人称射击游戏，自主运行约 4 天、42 轮规划/开发/验证，同时产出海报、演示文稿和网站。
- Raven RSI：对 nanochat 做预训练优化，7 轮共 172 次训练无崩溃，在相同的 20 分钟单 GPU 预算下 val_bpb 降低 5.8%；溃坝模拟的超调降低三个数量级；FEA 极限载荷搜索 8 轮二分完成。
- 产品发布套件：网页小游戏（Angry Raven）、16 页概览、中英海报，以及项目 README 本身。
这些是项目方的展示，价值在于说明"长时、多轮、可验证"的任务形态，而不是作为通用性能证明。
目录结构
仓库顶层包括 agents、benchmarks、evolver、experimental、raven、plugins-dist、skills、ui-tui、ui-web、docker、docs、schemas、rpc-schema 等目录，以及 AGENTS.md、CLAUDE.md、CONTEXT.md、CONTEXT-MAP.md、CONTRIBUTING.md、CHANGELOG.md、RELEASING.md 等文件。其中 experimental 对应前面提到的实验性能力，benchmarks 与 evolver 则是自我改进闭环的评测一侧。
局限与需要留意的点
- Curator 为实验性，且不在安装包内，自进化的完整能力需要从仓库源码使用。
- 性能数据多来自项目自述，需要自行验证在你的任务上的表现。
- 成本：多智能体长时运行会叠加模型调用，Oncall 这类通宵任务尤其要提前规划预算。
- 信任边界：一键安装脚本是 curl | bash 形式，生产环境建议先阅读脚本，或使用 Docker 方式隔离。
- 项目页没有提供配置详解和路线图，深度使用要依赖官方文档。
项目地址与资源
官方资源
- 🌟 GitHub: github.com/EverMind-AI…
- 📚 文档: evermind-ai.github.io/Raven/
- 🌐 官网: raven.evermind.ai
- 📄 论文: arXiv 2609.33439
- 🐛 Issue Tracker: github.com/EverMind-AI…
相关资源
- EverMind 生态：EverOS、EverMe、SkillCorpus、EvoAgentBench 等
- 集成项目：OpenClaw、Hermes Agent、DeepSeek Harness、Dify 插件
总结与展望
核心要点回顾
- 定位清晰：Raven 是 Host Agent，编排内置与第三方智能体，而不是又一个单体智能体。
- harness 可进化：Agent loop 拆成 Memory / Planning / Capability / Action 四个解耦模块，Evolver 与 Curator 在评测门控下改进它们。
- 生态完整：13 个预设第三方智能体、EverOS 记忆、SkillForge 技能检索、Proactivity 主动执行。
- 需要审慎看待：自进化核心（Curator）仍是实验性，性能数字需要自行验证。
适用人群
- 已经在用 Claude Code / Codex 等多个智能体、想把它们统一编排的开发者
- 研究 harness 工程、智能体自我改进（RSI）的研究者和工程师
- 需要长时无人值守任务（实验、监控、批量产出）的团队
学习建议
先用 Docker 或安装脚本跑起 raven web，选一个小任务观察 Host 如何调度智能体；再读技术报告和 evolver、benchmarks 目录，理解它如何用评测给自我改进设门槛。
探索 PrimeSkills —— 精选 AI agent 和技能工具，每一个都经过真实工作流验证。没有炒作，只有真正好用的工具。
访问我的个人主页，获取更多见解和有趣的产品。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Introduction
"The Harness of Harnesses."
This is the 245th article in the "One Open Source Project a Day" series. Today's project for you is Raven.
In the last year, Claude Code, Codex, OpenCode, Hermes Agent, such as coding / General AI agent, each with its own habitat (cycle, tools, memory, skill combination). The question follows: who will dispatch so many AI agents? Who decides which mission to give? Can you make yourself better?
Raven gives a clear position: it's a Host Agent that combines built-in AI agent and third party AI agent to perform complex tasks and, through modularization, makes harness itself an object that can be improved from time to time.
What will you learn?
- Raven's location: "Harness of Harnesses"
- Four built-in AI agents.
- Self-evolved while running: Agent loop to which of the four decomposition modules, Curator to what?
- Five core systems: Organization, Evolver, EverOS Memory, SkillForge, Practice
- How to install and start WebUI
- Who does it fit? Where is it limited?
Pre-knowledge
- Understanding the basic cycle of LLM Agent (planning tool call observation)
- AI agent with Claude Code / Codex etc.
- General impression of AI agent access protocol for MCP / ACP
Project background
Project Profile
Raven calls himself "Multi-AI agent Ecology Towards Recursive Self-improvement (RSI). It does three things:
- Hostage and organization: as Host Agent, coordinating multiple AI agentes, dependency and parallel.
- Self-improvement: The modularization architecture allows the iterative improvement of its own harness - proposes changes, evaluates, adopts validated improvements.
- Persistence memory: the memory layer is provided by EverOS and context can be kept between sessions.
Introduction by author/team
- Team: EverMind-AI
- Technical report: Raven: The Harness of Harnesses for Composable Industry, arXiv 2609.33439 (project page claims that the paper was awarded to Hugging Face Paper of the Day I, 2026-09-30)
- Ecology: EverMind also has EverOS (Memory), EverMe (CLI) and SkillCorpus, EverAlgo, HyperMem, MSA, EverMembench, EvoAgentBench, etc., research and evaluation projects, integrated with OpenClaw, Hermes Agent, DeepSeek Harness, Dify plugins
- Community: X, Discord, WeCom
- Project creation time: not specified on project page
Project data
- ⭐ GitHub Stars: 5.7k
- 🍴 Forks: 198
- 📝 Commits: 1,910
- 📄 License: Apache-2.0
- 🌐 🌐 ra ra
- Document: evermind-ai.github.io/Raven
Main functions
Core role
Raven was trying to solve the two questions: "How do you combine the multiple Q42s, how do you make it?" On the one hand, different AI agents are used as a combustible unit of capability to be centrally deployed by Host; on the other hand, key strategies in Agent Loop are removed so that they can be replaced, measured and evolved rather than written in a product.
Use scene
- Autonomous depth studies
- Structured reports of traceable sources for the Raven-Research export belt.
- AI agent software development
- Raven-Code is responsible for functional realization, debugging, re-engineering and data processing and analysis.
- Visual design outputs
- Raven-Design generates PPT, brands asset, charts, web interfaces.
- Unattended long assignment.
- Raven-Oncall does experiments, optimisation, continuous monitoring, which can last for hours or even overnight.
- Cross-AI agent collaboration
- AI agent has been accessed as a third party by Claude Code, Codex and others, organized by Raven.
Fast start.
The most cost-effective way is for your already existing AI agent to install it and give it the following:
Read https://evermind-ai.github.io/Raven/quick-start/ and follow it to install Raven, or to update it if it is already installed.
Manual installation:
# Linux / macOS / WSL2
curl -fsSL https://raven.evermind.ai/install.sh | bash
# Windows PowerShell
irm https://raven.evermind.ai/install.ps1 | iex
Docker Method (need Git and Docker Company):
git clone https://github.com/EverMind-AI/Raven.git
cd Raven
docker compose -f docker/docker-compose.yml up
# Then visit http://localhost:18793
Start WebUI:
raven web # Run and open browser
♪ Stop
WebUI supports dialogue, multi-AI agent collaboration, working area management, viewing documents and outputs, browsing memory and skills.
Install scripts to run directly for editable installation; publish weel is installed through conduit execution. This will allow you to choose between each AI agent product: using its modulation model (which requires a separate key) or using the LLM of Raven's own configuration.
Core characteristics
- Four internal AI agents.
- Research, Code, Design, Oncall, covering research, development, design, long-term automation.
- 13 preset third party AI agent
- API-compatible access through ACP, CLI or Open AI: Claude Code, Codex, OpenCode, Hermes Agent, OpenClaw, Mirothinker, GitHub Copilot, Qwen Code, CodeBudy, Qoder, Grok Build, Kimi Code, Pi.
- Evolver: Assessing the dynamics of evolution
- Failed diagnosis and improved candidate testing, with only better changes than baseline.
- Curator
- Continuously rewrite the four policy modules of Agent Loop.
- EverOS Cross-Memorial
- Preservation of user context, experience and knowledge.
- SkillForge
- Retrieving skills from local banks, EverOS memory and SkillHub (skill Hub is 114,190 on the project page).
- Proactivity
- Incident monitoring and regular implementation, active alerting and follow-up.
- Persona
- Description of the assistant, Curator, assembles a main role and specialists.
Project strengths
The project page does not provide an official comparison with similar projects. The following is a qualitative comparison based on my location, for reference purposes only:
Why should I care?
- It's turning "harness is something that can be improved by engineering" into a product, not a paper.
- It doesn't ask you to give up Claude Code/ Codex, but to put them in the set.
- There are thresholds for improvement: the project page emphasizes non-installation, failure or return to work before validation, and the signal usually strips off the reference answer.
Detailed Project Profile
Corporate structure: Host + Workers + 5 systems
The Raven core system can be summarized in five pieces:
Inline AI agent
These figures are derived from the project itself and have not been reproduced by me independently, and the reader suggested to return to the paper and benchmarks directory reconciliation before citing it.
Shared Harness by Raven Evolver. Evolver is a stand-alone tool using Raven in a library to assess candidate changes against benchmarks.
Four decoupling modules for Agent Loop
That's the most interesting point in Raven's design. It splits the Agent Loop into four policy modules:
- Memoory: Remember what, how?
- Planning: How to dismantle the mission.
- Capability: What tools, skills
- Action: How to execute and judge
After the decoupling of four pieces, each block can be replaced and evaluated separately. Curator (experimentary, with warehouse publish, not in the installation package) will continuously rewrite the perimeter to replace tools, skills, processes and judgement logic.
This is in line with the line of thought that we have repeatedly emphasized in the series of assessments: to make a system sustainable, first, it needs to have clear boundaries for its variable parts, and second, credible assessments to determine whether or not to change it. Raven ' s quality assurance rules (non-installation before validation, failure to return to work, signal stripping off reference answers) are essentially about adding door control to self-improvement.
Presentation case
Three cases were given on the project page:
- THRESHOLD: A Godot 4 first person shooting game, which operates autonomously for approximately 4 days, 42 rounds of planning/development/certification, and produces posters, presentations and websites.
- Raven RSI: Pre-training optimized for nanochat seven rounds of 172 training without collapse, with a 5.8 per cent reduction in val bpb under the same 20-minute GPU budget; three orders of magnitude reduction in over-alignment emulation; FEA maximum load search completed in two parts of eight.
- Product publish package: Angry Raven, 16 page overview, Chinese and English posters, and project README itself.
These are project parties' exhibits, and their value is to describe "long, multi-wheel, verifiable" mission patterns, rather than as generic performance evidence.
Contents structure
The top floors of the warehouse include catalogues of agents, benchmarks, evolver, executive, reven, plugins-dist, skills, ui-tui, ui-web, docker, docs, schemas, rpc-schema, and documents AGENTS.md, CLAUDE.md, CONTEXT-MAP.md, CONTRIBUBUTING.md, CHANGELOG.md, RELEASING.md. Of these, experimental capabilities correspond to the experimental capabilities mentioned earlier, and benchmarks and evolver are on the side of self-improvement.
Limitations and points that need attention
- To be experimental and not included in the installation package, the complete capability of self-evolution needs to be used from the warehouse source code.
- Performance data are mostly derived from project descriptions and require self-validation of performance in your mission.
- Cost: Multiple AI agent hours of operation overlap with model calls, and oncall, in particular, pre-planned budgets for such overnight missions.
- Trusted boundaries: One key is installed in the form of a curl |bash, and the production environment suggests reading the script first, or using the Docker method of isolation.
- The project page does not provide a detailed configuration and road map and relies on official documents for in-depth use.
Project address and resources
Official resources
- 🌟 GitHub: github.com/EverMind-AI…
- Document: evermind-ai.github.io/Raven/
- 🌐 🌐 ra ra
Thesis: arXiv 2609.33439
- 🐛 Issue Tracker: github.com/EverMind-AI…
Related resources
- EverMind Ecology: EverOS, EverMe, SkyllCorpus, EvoAgentBench, etc.
- Integrated projects: OpenClaw, Hermes Agent, DeepSeek Harness, Diffy Plugin
Summary and outlook
Review of core elements
- Clear location: Raven is Host Agent, organized with third party AI agent, not another single AI agent.
- Evolution: Agent looping into four decorative modules of Memoory / Planning / Capitalization / Action, which Evolver and Curator refines under door control.
- Ecological integrity: 13 pre-set third parties AI agent, EverOS Memory, SkyllForge Skills Retrieval, Proactivity.
- Need to be viewed with caution: the Curator core is still experimental and performance numbers need to be self-validated.
Population applicable
- Are already using multiple AI agent developers like Claude Code / Codex who want to organize them together
- Researcher and engineering for harness engineering, AI agent self-improvement (RSI) Division
- Teams with long-term no-man assignments (experiment, monitoring, mass output)
Learning advice
Start running with Docker or install scripts, pick a small mission watch host how to move AI agent; read the technical report and the evalver, benchmarks directory to understand how it sets the threshold for self-improvement.
Explore PrimeSkills - Select AI angent and skill tools, each validated by real workflow. No tricks, only real good tools.
Visit my home page to get more insights and interesting products.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.