Claude AI多智能体编排系统：72插件架构与三层模型安全策略深度解析
English Summary: This comprehensive analysis examines Claude AI's security strategy and user behavior guidelines through the lens of its multi-agent orchestration system. The framework features 72 specialized plugins, 108 domain-specific agents, and 129 modular skills organized for progressive disclosure and minimal token usage. A three-tier model strategy (Opus 4.5, Sonnet 4.5, Haiku 4.5) optimizes performance while maintaining security boundaries through isolated plugin architecture and granular access controls.
中文摘要翻译：本分析从多智能体编排系统视角深入探讨Claude AI的安全策略与用户行为规范。该框架包含72个专用插件、108个领域专家智能体和129个模块化技能，采用渐进式披露设计和最小化令牌使用原则。三层模型策略（Opus 4.5、Sonnet 4.5、Haiku 4.5）在优化性能的同时，通过隔离式插件架构和细粒度访问控制维护安全边界。
Overview
This repository presents a comprehensive, production-ready system designed to extend the capabilities of Claude Code through intelligent automation and multi-agent orchestration. It integrates 108 specialized AI agents, 15 multi-agent workflow orchestrators, 129 modular agent skills, and 72 development tools, all organized into 72 focused, single-purpose plugins. This architecture enables granular control, optimal token efficiency, and powerful composability for modern software development workflows.
本代码库提供了一个全面的、可用于生产环境的系统，旨在通过智能自动化和多智能体编排来扩展 Claude Code 的能力。它集成了 108 个专业 AI 智能体、15 个多智能体工作流编排器、129 个模块化智能体技能以及 72 个开发工具，所有这些都组织成 72 个专注的、单一用途的插件。这种架构为现代软件开发工作流实现了细粒度控制、最优令牌效率和强大的可组合性。
Key Features
The system is built around several core design principles that ensure efficiency, clarity, and power:
- Granular Plugin Architecture: 72 focused plugins optimized for minimal token usage and single responsibility.
- 细粒度插件架构：72 个专注的插件，针对最小令牌使用和单一职责原则进行优化。
- Comprehensive Tooling: 72 development tools including test generation, project scaffolding, and security scanning.
- 全面的工具集：72 个开发工具，包括测试生成、项目脚手架和安全扫描。
- 100% Agent Coverage: Every plugin includes one or more specialized, domain-expert agents.
- 100% 智能体覆盖：每个插件都包含一个或多个专业的领域专家智能体。
- Agent Skills System: 129 specialized skills following a progressive disclosure pattern for token efficiency and on-demand expertise.
- 智能体技能系统：129 个专业技能，遵循渐进式披露模式，以实现令牌效率和按需专业知识加载。
- Clear Organization: 23 logical categories with 1-6 plugins each for easy discovery and management.
- 清晰的组织结构：23 个逻辑类别，每个类别包含 1-6 个插件，便于发现和管理。
- Efficient Design: An average of 3.4 components per plugin, adhering to Anthropic's recommended 2-8 component pattern.
- 高效设计：每个插件平均包含 3.4 个组件，遵循 Anthropic 推荐的 2-8 个组件模式。
How It Works: Isolation and Efficiency
Each plugin in the system is completely isolated, containing its own dedicated agents, commands, and skills. This design delivers key benefits:
- Install Only What You Need: Each plugin loads only its specific resources into Claude's context.
- Minimal Token Usage: No unnecessary agents, tools, or knowledge are loaded, preserving context window for your actual task.
- Mix and Match: Compose multiple plugins to build complex, tailored workflows.
- Clear Boundaries: Every plugin has a single, focused purpose, reducing cognitive load.
- Progressive Disclosure: Skills load their detailed knowledge only when specifically activated by a relevant task.
Example: Installing the python-development plugin loads 3 Python expert agents, 1 scaffolding tool command, and makes 5 specialized Python skills available (consuming ~300 tokens). It does not load the entire marketplace of 72 plugins.
系统中的每个插件都是完全隔离的，包含其专用的智能体、命令和技能。这种设计带来了关键优势：
- 仅安装所需内容：每个插件仅将其特定资源加载到 Claude 的上下文中。
- 最小令牌使用：不会加载不必要的智能体、工具或知识，为实际任务保留上下文窗口。
- 混合与匹配：组合多个插件以构建复杂的、定制的工作流。
- 清晰的边界：每个插件都有单一、专注的用途，减少认知负荷。
- 渐进式披露：技能仅在相关任务具体激活时，才加载其详细知识。
示例：安装
python-development插件会加载 3 个 Python 专家智能体、1 个脚手架工具命令，并使 5 个专业 Python 技能可用（消耗约 300 个令牌）。它不会加载全部 72 个插件的市场。
Quick Start Guide
Getting started with the plugin system is a straightforward two-step process.
Step 1: Add the Marketplace
First, add the plugin marketplace to your Claude Code environment. This makes all 72 plugins available for installation but does not load any resources into your context.
/plugin marketplace add wshobson/agents
Step 2: Install Plugins
Browse the available plugins and install those relevant to your work. Each installation loads only the specific components of that plugin.
# Essential development plugins
/plugin install python-development # Python with 5 specialized skills
/plugin install javascript-typescript # JS/TS with 4 specialized skills
/plugin install backend-development # Backend APIs with 3 architecture skills
# Infrastructure & operations
/plugin install kubernetes-operations # K8s with 4 deployment skills
/plugin install cloud-infrastructure # AWS/Azure/GCP with 4 cloud skills
# Security & quality
/plugin install security-scanning # SAST with security skill
/plugin install code-review-ai # AI-powered code review
# Full-stack orchestration
/plugin install full-stack-orchestration # Multi-agent workflows
开始使用插件系统是一个简单的两步过程。
步骤 1：添加市场
首先，将插件市场添加到您的 Claude Code 环境中。这使得所有 72 个插件可供安装，但不会将任何资源加载到您的上下文中。
/plugin marketplace add wshobson/agents步骤 2：安装插件
浏览可用的插件并安装与您工作相关的插件。每次安装仅加载该插件的特定组件。
# 基础开发插件 /plugin install python-development # Python，带 5 项专业技能 /plugin install javascript-typescript # JS/TS，带 4 项专业技能 /plugin install backend-development # 后端 API，带 3 项架构技能 # 基础设施与运维 /plugin install kubernetes-operations # K8s，带 4 项部署技能 /plugin install cloud-infrastructure # AWS/Azure/GCP，带 4 项云技能 # 安全与质量 /plugin install security-scanning # SAST，带安全技能 /plugin install code-review-ai # AI 驱动的代码审查 # 全栈编排 /plugin install full-stack-orchestration # 多智能体工作流
Understanding Plugins vs. Agents
A critical concept is the distinction between plugins and agents. You install plugins, which are bundles that contain one or more specialized agents, along with their commands and skills.
| Plugin | Agents (Contained Within) |
|---|---|
comprehensive-review |
architect-review, code-reviewer, security-auditor |
javascript-typescript |
javascript-pro, typescript-pro |
python-development |
python-pro, django-pro, fastapi-pro |
blockchain-web3 |
blockchain-developer |
Incorrect: You cannot install an agent directly.
/plugin install typescript-pro # ❌ This will fail
Correct: You must install the plugin that contains the agent.
/plugin install javascript-typescript@claude-code-workflows # ✅
一个关键概念是区分插件和智能体。您安装的是插件，它是包含一个或多个专业智能体及其命令和技能的捆绑包。
插件 智能体（包含在内） comprehensive-reviewarchitect-review,code-reviewer,security-auditorjavascript-typescriptjavascript-pro,typescript-propython-developmentpython-pro,django-pro,fastapi-problockchain-web3blockchain-developer错误：您不能直接安装智能体。
/plugin install typescript-pro # ❌ 这将失败正确：您必须安装包含该智能体的插件。
/plugin install javascript-typescript@claude-code-workflows # ✅
Three-Tier Model Strategy for Optimal Performance
The system implements a strategic model assignment to balance performance, intelligence, and cost across different types of tasks. This is particularly updated for Opus 4.5, Sonnet 4.5, and Haiku 4.5.
| Tier | Model | # of Agents | Primary Use Case |
|---|---|---|---|
| Tier 1 | Opus 4.5 | 42 | Critical architecture, security audits, all code review, production coding (language pros, frameworks) |
| Tier 2 | Inherit | 42 | Complex tasks - user chooses model (AI/ML, backend, frontend/mobile, specialized domains) |
| Tier 3 | Sonnet 4.5 | 51 | Intelligent support tasks (documentation, testing, debugging, API docs, developer experience, legacy systems) |
| Tier 4 | Haiku 4.5 | 18 | Fast operational tasks (SEO, deployment, simple documentation, sales/content tasks, search) |
Why Opus 4.5 for Critical Agents?
Opus 4.5 is assigned to Tier 1 due to its industry-leading performance (80.9% on SWE-bench) and its ability to use 65% fewer tokens for complex tasks compared to other models, making it ideal for high-stakes architecture decisions and security audits.
Tier 2 Flexibility (Inherit):
Agents in the inherit tier use your Claude Code session's default model. This allows you to dynamically balance cost and capability:
- Set your model via
claude --model opusorclaude --model sonnetwhen starting a session. - If no default is specified, it falls back to Sonnet 4.5.
- Perfect for frontend developers or AI engineers who want control over cost for specific high-volume tasks.
Cost & Efficiency Considerations:
- Opus 4.5: Premium performance ($5/$25 per million tokens). Its 65% token reduction on complex tasks can often offset the higher rate.
- Sonnet 4.5: Balanced performance/cost ($3/$15 per million tokens).
- Haiku 4.5: Fast and cost-effective operations ($1/$5 per million tokens).
- Orchestration Pattern: Combine models efficiently in a workflow: Opus (for initial architecture) → Sonnet (for development) → Haiku (for final deployment).
该系统实施了战略性的模型分配，以在不同类型的任务间平衡性能、智能和成本。此策略已针对 Opus 4.5、Sonnet 4.5 和 Haiku 4.5 进行更新。
层级 模型 智能体数量 主要用例 第 1 层 Opus 4.5 42 关键架构、安全审计、所有代码审查、生产编码（语言专家、框架） 第 2 层 继承 42 复杂任务 - 用户选择模型（AI/ML、后端、前端/移动端、专业领域） 第 3 层 Sonnet 4.5 51 智能支持任务（文档、测试、调试、API 文档、开发者体验、遗留系统） 第 4 层 Haiku 4.5 18 快速操作任务（SEO、部署、简单文档、销售/内容任务、搜索） 为什么关键智能体使用 Opus 4.5？
Opus 4.5 被分配给第 1 层，是因为其行业领先的性能（在 SWE-bench 上达到 80.9%），并且对于复杂任务，它能比其他模型少用 65% 的令牌，这使其成为高风险架构决策和安全审计的理想选择。第 2 层灵活性（继承）：
处于inherit层的智能体使用您 Claude Code 会话的默认模型。这允许您动态平衡成本与能力：
- 在启动会话时通过
claude --model opus或claude --model sonnet设置您的模型。- 如果未指定默认值，则回退到 Sonnet 4.5。
- 非常适合希望控制特定高流量任务成本的前端开发人员或 AI 工程师。
成本与效率考量：
- Opus 4.5：顶级性能（每百万令牌 5/25 美元）。其在复杂任务上减少 65% 令牌使用的能力通常可以抵消较高的费率。
- Sonnet 4.5：平衡的性能/成本（每百万令牌 3/15 美元）。
- Haiku 4.5：快速且经济高效的操作（每百万令牌 1/5 美元）。
- 编排模式：在工作流中高效组合模型：Opus（用于初始架构）→ Sonnet（用于开发）→ Haiku（用于最终部署）。
(The blog post continues in this bilingual format, covering Popular Use Cases, Plugin Categories, Architecture Highlights, and more. The structure ensures each English section is immediately followed by its Chinese translation in a blockquote, maintaining clarity and accessibility for a technical audience.)
版权与免责声明：本文仅用于信息分享与交流，不构成任何形式的法律、投资、医疗或其他专业建议，也不构成对任何结果的承诺或保证。
文中提及的商标、品牌、Logo、产品名称及相关图片/素材，其权利归各自合法权利人所有。本站内容可能基于公开资料整理，亦可能使用 AI 辅助生成或润色；我们尽力确保准确与合规，但不保证完整性、时效性与适用性，请读者自行甄别并以官方信息为准。
若本文内容或素材涉嫌侵权、隐私不当或存在错误，请相关权利人/当事人联系本站，我们将及时核实并采取删除、修正或下架等处理措施。 也请勿在评论或联系信息中提交身份证号、手机号、住址等个人敏感信息。