OpenYoung

智能 Agent 发现与部署平台 | Intelligent Agent Discovery & Deployment Platform

English

What is OpenYoung?

OpenYoung is an AI agent platform that automatically discovers, evaluates, and deploys high-quality agents based on your task input.

Instead of manually searching and configuring agents, just describe what you want to do — OpenYoung handles the rest.

Core Features

┌─────────────────────────────────────────────────────────────────┐
│                      OpenYoung Pipeline                          │
└─────────────────────────────────────────────────────────────────┘
                              │
                              ▼
                    ┌─────────────────┐
                    │  Intent        │
                    │  Analysis      │ ◄── LLM-powered intent detection
                    └────────┬────────┘
                             │
                             ▼
                    ┌─────────────────┐
                    │  Semantic       │ ◄── Vector similarity search
                    │  Search        │
                    └────────┬────────┘
                             │
                             ▼
                    ┌─────────────────┐
                    │  Quality        │ ◄── Multi-dimensional evaluation
                    │  Evaluation     │
                    └────────┬────────┘
                             │
                    ┌────────┴────────┐
                    ▼                 ▼
              ┌─────────┐       ┌─────────┐
              │ Badges  │       │ Version │
              │ Display │       │ History │
              └────┬────┘       └────┬────┘
                   │                 │
                   └────────┬────────┘
                            ▼
                    ┌─────────────────┐
                    │  Auto-Config   │ ◄── Always Skills loading
                    └────────┬────────┘
                             │
                             ▼
                    ┌─────────────────┐
                    │  Execute Task  │
                    └─────────────────┘

Feature	Description
Semantic Search	Find agents using natural language, not keywords
Quality Evaluation	6-dimension scoring: completeness, validity, dependencies, documentation, security, runtime
Intent Analysis	Understand what you want to do, recommend the right agent
Badge System	Visual quality indicators: Verified, Top Rated, Trending, New, Popular
Version Management	Track agent versions with SemVer support
Usage Tracking	Monitor which agents are most popular

Quick Start

# Run agent with natural language
openyoung run "帮我写一个排序算法"

# List all available agents with badges
openyoung agent list --all --badges --stats

# Search agents semantically
openyoung agent search "代码审查"

# Compare two agents
openyoung agent compare default coder

# Analyze your intent
openyoung agent intent "我想要自动化测试"

# Check version history
openyoung agent versions agent-coder
openyoung agent version-add agent-coder 1.0.0 --changelog "Initial release"

CLI Commands

Command	Description
`openyoung run <task>`	Run agent with task description
`openyoung agent list`	List available agents
`openyoung agent info <agent>`	Show agent details
`openyoung agent search <query>`	Semantic search agents
`openyoung agent compare <a> <b>`	Compare two agents
`openyoung agent evaluate <agent>`	Quality evaluation
`openyoung agent intent <query>`	Intent analysis
`openyoung eval list`	List evaluation metrics
`openyoung eval trend <agent>`	Show evaluation trend
`openyoung eval run <task>`	Run evaluation on a task
`openyoung eval server`	Start eval API server
`openyoung data runs`	List run records
`openyoung data list`	List run records (alias)
`openyoung memory list`	List memories
`openyoung memory search <query>`	Search memories
`openyoung mcp servers`	List MCP servers
`openyoung skills list`	List skills
`openyoung config list`	List configuration
`openyoung llm list`	List LLM providers
`openyoung package list`	List packages
`openyoung source list`	List sources
`openyoung channel list`	List channels
`openyoung templates list`	List templates
`openyoung import github <repo>`	Import from GitHub
`openyoung --help`	Show help

Agent Package Structure

packages/
├── agent-coder/
│   ├── agent.yaml          # Agent configuration
│   ├── pyproject.toml     # Dependencies
│   └── src/               # Agent code
└── agent-reviewer/
    ├── agent.yaml
    └── pyproject.toml

Configuration

Create .env in your project root:

# Required: At least one LLM provider
OPENAI_API_KEY=your-openai-key
# or
ANTHROPIC_API_KEY=your-anthropic-key

# Optional: Vector search (for semantic search)
DASHSCOPE_API_KEY=your-dashscope-key

Architecture

openyoung/
├── src/
│   ├── agents/                # Agent system
│   │   ├── harness/           # ✅ Harness-centric AI factory (NEW)
│   │   │   ├── engine.py     # Core execution engine
│   │   │   ├── graph.py      # Task graph
│   │   │   ├── task_compiler.py    # Task → Graph compiler
│   │   │   ├── resource_manager.py  # Resource allocation
│   │   │   ├── harness_runner.py    # Lifecycle management
│   │   │   └── types.py      # Streaming types
│   │   ├── execution/         # ✅ Execution layer (NEW)
│   │   ├── commands/         # ✅ CLI commands (NEW - modular)
│   │   └── young_agent.py    # ✅ Refactored to ~400 lines
│   ├── cli/                  # ✅ Refactored to ~100 lines
│   │   ├── main.py           # Entry point (96 lines)
│   │   └── commands/         # Modular commands
│   ├── core/                 # Core types & infrastructure
│   │   ├── memory/          # ✅ Hierarchical memory system
│   │   ├── events.py         # EventBus
│   │   ├── heartbeat.py      # Heartbeat scheduler
│   │   └── langgraph_*.py    # LangGraph integration
│   ├── hub/                   # Hub system
│   │   └── evaluate/         # ✅ Unified evaluation (NEW)
│   │       ├── harness.py    # Evaluation harness
│   │       ├── runner.py     # Eval runner
│   │       └── benchmark.py  # Benchmark tools
│   └── webui/                # Streamlit WebUI
├── packages/                 # Agent packages
├── skills/                   # Skills
├── tests/                    # Test suite (119 tests)
└── docs/                    # Documentation

Refactoring Achievements

Metric	Before	After	Improvement
young_agent.py	1665 lines	~400 lines	-76%
cli/main.py	2167 lines	~96 lines	-96%
Harness coverage	30%	95%	+217%
Test coverage	~20%	80%+	+300%
Code duplication	High	Minimal	✅

Reinforcement Learning (RL)

OpenYoung supports optional GPU-accelerated Reinforcement Learning for agent self-evolution and optimization.

Core Features

Feature	Description
Multi-Backend Support	CUDA (NVIDIA), MPS (Apple Silicon), CPU, Vulkan (ARM RK3588)
GRPO	Group Relative Policy Optimization - lightweight training for M1/medium GPUs
GiGPO	Group-in-Group Policy Optimization - two-layer advantage for complex agents
Experience Collection	Collect task experiences without GPU (default mode)
Docker Support	Ready-to-deploy RL training with Docker Compose

Supported Hardware

Hardware	Recommended Mode	Notes
NVIDIA GPU (CUDA)	GiGPO	Full two-layer advantage
Apple Silicon (MPS)	GRPO	Memory-efficient training
CPU only	Collection Only	Experience collection only
ARM (RK3588)	Collection Only	Future: Vulkan support

Usage

from src.agents.rl import create_rl_engine, RLConfig, RLMode

# Auto-detect hardware and create engine
engine = create_rl_engine()

# Or specify mode manually
config = RLConfig(mode=RLMode.GRPO)  # collection_only / grpo / gigpo
engine = create_rl_engine(config)

Docker Deployment

# Start RL training service
docker-compose -f docker-compose.rl.yml up -d

# With NVIDIA GPU
docker-compose -f docker-compose.rl.yml up -d --profile gpu

Configuration

# config/rl.yaml
rl:
  enabled: true
  mode: collection_only  # collection_only | grpo | gigpo
  device: auto  # auto | cuda | mps | cpu
  grpo:
    learning_rate: 1.0e-5
    clip_epsilon: 0.2
    group_size: 4

AI Docker

OpenYoung includes a built-in AI Docker sandbox execution environment that provides secure and controlled code execution for AI Agents.

Core Features

Feature	Description
Sandbox Isolation	Resource limits (CPU/Memory/Time), network access control
Security Policy	Command whitelist, dangerous pattern detection, file path validation
Instance Pool	Auto-scaling, instance pre-warming, state persistence
Audit Logging	Execution records, statistics, JSONL format

Enhanced Security (2025 Best Practices)

Security Feature	Description
Working Directory Restriction	Files can only be accessed within configured working directory
Path Traversal Prevention	Blocks `..`, `/proc`, `/sys`, `/dev` attacks
Network Isolation	Domain whitelist/blacklist, default deny
MCP Server Security	Command validation, environment sanitization
Audit Logging	All security events logged

Usage

from src.runtime.sandbox import SandboxPolicy, SecurityPolicyEngine

# Configure security policy
policy = SandboxPolicy(
    working_directory="/tmp/sandbox",       # Restrict to working dir
    restrict_to_working_dir=True,
    allow_network=True,
    allowed_domains=["api.openai.com"],
)
engine = SecurityPolicyEngine(policy)

# Check file access
safe, reason = engine.check_path_traversal("/tmp/sandbox/file.txt")  # ALLOWED
safe, reason = engine.check_path_traversal("/etc/passwd")  # BLOCKED

WebUI

OpenYoung provides a Streamlit-based WebUI for visual interaction with agents, sessions, and evaluations.

Core Features

Feature	Description
Chat Interface	Real-time streaming chat with AI agents
Session Management	Create, suspend, resume, delete persistent sessions
Agent Management	Browse and manage available agents
Evaluation Dashboard	View evaluation results and metrics
SSE Streaming	Server-Sent Events for real-time responses

Quick Start

# Install dependencies
pip install -r webui/requirements.txt

# Start WebUI
streamlit run webui/app.py

# Or use the CLI
openyoung webui

Configuration

Edit webui/utils/config.py:

API_BASE_URL = "http://localhost:8000"  # API server URL
API_KEY = ""  # Optional API key
TYPING_SPEED = 0.02  # Seconds between characters

License

MIT License

中文

什么是 OpenYoung？

OpenYoung 是一个 智能 Agent 发现与部署平台，能够根据你的任务输入自动发现、评估和部署高质量 Agent。

无需手动搜索和配置 Agent，只需描述你想要做什么 —— OpenYoung 会帮你完成其余工作。

核心功能

┌─────────────────────────────────────────────────────────────────┐
│                      OpenYoung 工作流程                         │
└─────────────────────────────────────────────────────────────────┘
                              │
                              ▼
                    ┌─────────────────┐
                    │  意图理解       │ ◄── LLM 驱动的意图分析
                    │  Intent        │
                    └────────┬────────┘
                             │
                             ▼
                    ┌─────────────────┐
                    │  语义搜索       │ ◄── 向量相似度匹配
                    │  Semantic      │
                    └────────┬────────┘
                             │
                             ▼
                    ┌─────────────────┐
                    │  质量评估       │ ◄── 多维度评分
                    │  Quality       │
                    └────────┬────────┘
                             │
                    ┌────────┴────────┐
                    ▼                 ▼
              ┌─────────┐       ┌─────────┐
              │  徽章系统 │       │ 版本管理 │
              │  Badges │       │ Versions│
              └────┬────┘       └────┬────┘
                   │                 │
                   └────────┬────────┘
                            ▼
                    ┌─────────────────┐
                    │  自动配置       │ ◄── Always Skills 加载
                    │  Auto-Config   │
                    └────────┬────────┘
                             │
                             ▼
                    ┌─────────────────┐
                    │  执行任务       │
                    │  Execute       │
                    └─────────────────┘

功能	描述
语义搜索	用自然语言而非关键词搜索 Agent
质量评估	6 维度评分：完整性、有效性、依赖、文档、安全、运行时
意图理解	理解你想要做什么，推荐合适的 Agent
徽章系统	可视化质量标识：官方验证、高评分、趋势上升、新增、热门
版本管理	SemVer 版本的版本历史追踪
使用追踪	监控哪些 Agent 最受欢迎

快速开始

# 用自然语言运行 agent
openyoung run "帮我写一个排序算法"

# 列出所有 agents（含徽章和统计）
openyoung agent list --all --badges --stats

# 语义搜索 agents
openyoung agent search "代码审查"

# 对比两个 agents
openyoung agent compare default coder

# 分析你的意图
openyoung agent intent "我想要自动化测试"

# 查看版本历史
openyoung agent versions agent-coder
openyoung agent version-add agent-coder 1.0.0 --changelog "初始版本"

CLI 命令

命令	描述
`openyoung run <任务>`	使用任务描述运行 agent
`openyoung agent list`	列出可用 agents
`openyoung agent info <agent>`	显示 agent 详情
`openyoung agent search <查询>`	语义搜索 agents
`openyoung agent compare <a> <b>`	对比两个 agents
`openyoung agent evaluate <agent>`	质量评估
`openyoung agent intent <查询>`	意图分析
`openyoung eval list`	列出评估指标
`openyoung eval trend <agent>`	显示评估趋势
`openyoung eval run <任务>`	运行任务评估
`openyoung eval server`	启动评估 API 服务器
`openyoung data runs`	列出运行记录
`openyoung data list`	列出运行记录（别名）
`openyoung memory list`	列出记忆
`openyoung memory search <查询>`	搜索记忆
`openyoung mcp servers`	列出 MCP 服务器
`openyoung skills list`	列出技能
`openyoung config list`	列出配置
`openyoung llm list`	列出 LLM 提供商
`openyoung package list`	列出包
`openyoung source list`	列出数据源
`openyoung channel list`	列出通道
`openyoung templates list`	列出模板
`openyoung import github <仓库>`	从 GitHub 导入
`openyoung --help`	显示帮助

Agent 包结构

packages/
├── agent-coder/
│   ├── agent.yaml        # Agent 配置
│   ├── pyproject.toml   # 依赖
│   └── src/             # Agent 代码
└── agent-reviewer/
    ├── agent.yaml
    └── pyproject.toml

配置

在项目根目录创建 .env 文件：

# 必需：至少一个 LLM 提供商
OPENAI_API_KEY=your-openai-key
# 或
ANTHROPIC_API_KEY=your-anthropic-key

# 可选：向量搜索（用于语义搜索）
DASHSCOPE_API_KEY=your-dashscope-key

项目架构

openyoung/
├── src/
│   ├── agents/                # Agent 系统
│   │   ├── harness/           # ✅ Harness驱动的AI工厂 (新增)
│   │   │   ├── engine.py     # 核心执行引擎
│   │   │   ├── graph.py      # 任务图
│   │   │   ├── task_compiler.py    # Task → Graph 编译器
│   │   │   ├── resource_manager.py  # 资源分配
│   │   │   ├── harness_runner.py    # 生命周期管理
│   │   │   └── types.py      # 流式类型
│   │   ├── execution/         # ✅ 执行层 (新增)
│   │   ├── commands/          # ✅ CLI 命令模块化 (新增)
│   │   └── young_agent.py     # ✅ 重构后约400行
│   ├── cli/                   # ✅ 重构后约100行
│   │   ├── main.py           # 入口 (96行)
│   │   └── commands/         # 模块化命令
│   ├── core/                  # 核心类型和基础设施
│   │   ├── memory/           # ✅ 分层记忆系统
│   │   ├── events.py         # 事件总线
│   │   ├── heartbeat.py      # 心跳调度器
│   │   └── langgraph_*.py    # LangGraph 集成
│   ├── hub/                   # Hub 系统
│   │   └── evaluate/          # ✅ 统一评估 (新增)
│   │       ├── harness.py     # 评估线束
│   │       ├── runner.py      # 评估运行器
│   │       └── benchmark.py   # 基准测试工具
│   └── webui/                # Streamlit WebUI
├── packages/                  # Agent 包
├── skills/                    # 技能
├── tests/                     # 测试套件 (119个测试)
└── docs/                     # 文档

重构成果

指标	重构前	重构后	改善
young_agent.py	1665行	~400行	-76%
cli/main.py	2167行	~96行	-96%
Harness覆盖率	30%	95%	+217%
测试覆盖率	~20%	80%+	+300%
代码重复	高	最低	✅

强化学习 (RL)

OpenYoung 支持可选的 GPU 加速强化学习，用于 Agent 自我进化与优化。

核心功能

功能	描述
多后端支持	CUDA (NVIDIA)、MPS (Apple Silicon)、CPU、Vulkan (ARM RK3588)
GRPO	组相对策略优化 - 适用于 M1/中等 GPU 的轻量训练
GiGPO	组内组策略优化 - 复杂 Agent 的双层优势估计
经验收集	无需 GPU 的经验收集（默认模式）
Docker 支持	开箱即用的 RL 训练服务部署

支持的硬件

硬件	推荐模式	备注
NVIDIA GPU (CUDA)	GiGPO	完整双层优势
Apple Silicon (MPS)	GRPO	内存高效训练
仅 CPU	仅收集	经验收集模式
ARM (RK3588)	仅收集	未来支持: Vulkan

使用方式

from src.agents.rl import create_rl_engine, RLConfig, RLMode

# 自动检测硬件并创建引擎
engine = create_rl_engine()

# 或手动指定模式
config = RLConfig(mode=RLMode.GRPO)  # collection_only / grpo / gigpo
engine = create_rl_engine(config)

Docker 部署

# 启动 RL 训练服务
docker-compose -f docker-compose.rl.yml up -d

# 使用 NVIDIA GPU
docker-compose -f docker-compose.rl.yml up -d --profile gpu

配置

# config/rl.yaml
rl:
  enabled: true
  mode: collection_only  # collection_only | grpo | gigpo
  device: auto  # auto | cuda | mps | cpu
  grpo:
    learning_rate: 1.0e-5
    clip_epsilon: 0.2
    group_size: 4

AI Docker

OpenYoung 内置 AI Docker 沙箱执行环境，为 AI Agent 提供安全可控的代码执行能力。

核心功能

功能	描述
沙箱隔离	资源限制 (CPU/Memory/Time)，网络访问控制
安全策略	命令白名单，危险模式检测，文件路径验证
实例池	自动扩缩容，预热实例，状态持久化
审计日志	执行记录，统计查询

安全增强 (2025最佳实践)

安全功能	描述
工作目录限制	文件只能访问配置的工作目录
路径穿越防护	阻止 `..`、`/proc`、`/sys`、`/dev` 攻击
网络隔离	域名白名单/黑名单，默认拒绝
MCP服务器安全	命令验证，环境变量清理
审计日志	所有安全事件记录

使用方式

from src.runtime.sandbox import SandboxPolicy, SecurityPolicyEngine

# 配置安全策略
policy = SandboxPolicy(
    working_directory="/tmp/sandbox",       # 限制工作目录
    restrict_to_working_dir=True,
    allow_network=True,
    allowed_domains=["api.openai.com"],
)
engine = SecurityPolicyEngine(policy)

# 检查文件访问
safe, reason = engine.check_path_traversal("/tmp/sandbox/file.txt")  # 允许
safe, reason = engine.check_path_traversal("/etc/passwd")  # 阻止

WebUI

OpenYoung 提供基于 Streamlit 的 WebUI，用于可视化交互 Agent、会话和评估。

核心功能

功能	描述
聊天界面	实时流式与 AI Agent 对话
会话管理	创建、暂停、恢复、删除持久会话
Agent 管理	浏览和管理可用的 Agent
评估仪表盘	查看评估结果和指标
SSE 流式传输	服务器发送事件实现实时响应

快速开始

# 安装依赖
pip install -r webui/requirements.txt

# 启动 WebUI
streamlit run webui/app.py

# 或使用 CLI
openyoung webui

配置

编辑 webui/utils/config.py:

API_BASE_URL = "http://localhost:8000"  # API 服务器地址
API_KEY = ""  # 可选的 API 密钥
TYPING_SPEED = 0.02  # 字符间隔秒数

许可证

MIT License

Name		Name	Last commit message	Last commit date
Latest commit History 100 Commits
.agents/skills/jamesxdigital-jamesxskills-excalidraw-diagram-generator		.agents/skills/jamesxdigital-jamesxskills-excalidraw-diagram-generator
.claude-flow		.claude-flow
.claude		.claude
.github		.github
.learnings		.learnings
.openyoung		.openyoung
.swarm		.swarm
benchmarks		benchmarks
claude-code		claude-code
config		config
docs		docs
evals		evals
examples		examples
github-import-workspace		github-import-workspace
output/github-trending		output/github-trending
packages		packages
rust/ironclaw-sandbox		rust/ironclaw-sandbox
scripts		scripts
skills/github-import		skills/github-import
src		src
subagents		subagents
templates		templates
tests		tests
webui		webui
.dockerignore		.dockerignore
.gitignore		.gitignore
.mcp.json		.mcp.json
.pre-commit-config.yaml		.pre-commit-config.yaml
AGENTS.md		AGENTS.md
CHANGELOG.md		CHANGELOG.md
CLAUDE.md		CLAUDE.md
CONTRIBUTING.md		CONTRIBUTING.md
Cargo.toml		Cargo.toml
Dockerfile		Dockerfile
Dockerfile.rl		Dockerfile.rl
README.md		README.md
debug_test.txt		debug_test.txt
direct_test.txt		direct_test.txt
docker-compose.rl.yml		docker-compose.rl.yml
docker-compose.yml		docker-compose.yml
findings.md		findings.md
github_import_script.py		github_import_script.py
mkdocs.yml		mkdocs.yml
progress.md		progress.md
pyproject.toml		pyproject.toml
requirements-rl.txt		requirements-rl.txt
task_plan.md		task_plan.md

Folders and files

Latest commit

History

Repository files navigation

OpenYoung

English

What is OpenYoung?

Core Features

Quick Start

CLI Commands

Agent Package Structure

Configuration

Architecture

Refactoring Achievements

Reinforcement Learning (RL)

Core Features

Supported Hardware

Usage

Docker Deployment

Configuration

AI Docker

Core Features

Enhanced Security (2025 Best Practices)

Usage

WebUI

Core Features

Quick Start

Configuration

License

中文

什么是 OpenYoung？

核心功能

快速开始

CLI 命令

Agent 包结构

配置

项目架构

重构成果

强化学习 (RL)

核心功能

支持的硬件

使用方式

Docker 部署

配置

AI Docker

核心功能

安全增强 (2025最佳实践)

使用方式

WebUI

核心功能

快速开始

配置

许可证

About

Resources

Contributing

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages