← Diagram Lab
Claude 官方架构图与代码图解
从 anthropic.com 与 claude.com 的 389 篇文章、1,591 张配图中，剔除评测/基准类图表，留下 420 张技术解释图，再人工筛出 78 张真正的架构、系统、代码与流程类图解。用来回答一个问题：他们到底怎么把系统画清楚。
1,591官方配图
420技术类图解
78本页精选
4可复用图式
六个结论
- 主力不是“插画”，而是带边界的系统图。最高频的成功图式是泳道/面板 + 嵌套容器，而不是装饰性插图。
- 颜色是语义，不是装饰。蓝=调度者/主流程，绿=子任务/允许，珊瑚=错误/拒绝/重点，米色=容器边界。跨文章高度一致。
- 一图只讲一个变量。四张多智能体图共用同一套方块语法，只替换拓扑；三张 workflow 图只替换支路。
- 最好的图会“标注边界”。User request、final report、allowed:443、unknown:80 这类标签直接写在箭头上，读者不用猜。
- 代码类图不截 IDE。他们画 SKILL.md 的结构、目录树、接口表；只有终端类内容才用截图。
- 质量不稳定。导出型泳道图（绿底绿字、箭头交叉）与过度简化的“一个方块一个圆”仍然存在，说明没有统一制图规范。
Agent 架构与多智能体协作
协作拓扑：orchestrator、agent teams、message bus、shared state 四种底座，以及 harness / SDK loop / generator–verifier。
28 张
Multi-agent coordination patterns: Five approaches and when to use them
最小可用语法：蓝色 orchestrator + 三个绿色 subagent。直角折线、无图例也能读懂；但回传箭头没有标签，方向语义靠猜。
Multi-agent coordination patterns: Five approaches and when to use them
同一套视觉语言换拓扑：lead 在中间，成员并行展开。
Multi-agent coordination patterns: Five approaches and when to use them
加一条橙色 message bus，用“通道”替代点对点连线，复杂协作立刻可读。
Multi-agent coordination patterns: Five approaches and when to use them
共享状态画成中心对象，所有 agent 读写同一块，比满天飞箭头清楚。
Multi-agent coordination patterns: Five approaches and when to use them
对比图：两种拓扑并排，标题即结论。
Multi-agent coordination patterns: Five approaches and when to use them
对比图：拓扑差异用橙色总线一眼区分。
Multi-agent coordination patterns: Five approaches and when to use them
对比图：状态归属是唯一变量。
Multi-agent coordination patterns: Five approaches and when to use them
对比图：message bus vs shared state，适合做“选型”页。
How we built our multi-agent research system
双面板：左侧真实对话，右侧系统架构，箭头标注 user request / final report——把“输入输出”钉死在边界上。
How we built our multi-agent research system
反面教材：泳道图直接导出，绿底绿框对比度低、字号过小、箭头交叉，读者需要放大才能读。
Building agents: When to fold subagents into your main agent loop
Building agents: When to fold subagents into your main agent loop
Building agents with Skills: Equipping agents for specialized work
Building agents with Skills: Equipping agents for specialized work
Building agents with Skills: Equipping agents for specialized work
珊瑚色描边 + 像素螃蟹图标，在严肃架构图里保留品牌温度。
Building agents with the Claude Agent SDK
SDK 主循环三色块：Gather context → Take action → Verify work，最经典的 agent loop 表达。
Lessons from Anthropic on building effective human-agent teams
树状任务分解：上层一个 agent，下层并行子任务，适合讲 fan-out。
Lessons from Anthropic on building effective human-agent teams
Building verification loops in Claude Code with skills
The evolution of agentic surfaces: building with Claude Managed Agents
The evolution of agentic surfaces: building with Claude Managed Agents
The evolution of agentic surfaces: building with Claude Managed Agents
The evolution of agentic surfaces: building with Claude Managed Agents
The evolution of agentic surfaces: building with Claude Managed Agents
Giving companies more control over their AI agents, with NVIDIA
How to prepare for AI-driven code modernization projects
Multi-agent coordination patterns: Five approaches and when to use them
Skills explained: How Skills compares to prompts, Projects, MCP, and subagents
Harness、工具与上下文管线
用泳道与边界表达权限、网络、上下文窗口、工具调用和沙箱——Anthropic 最成熟的一类图。
35 张
Code execution with MCP for efficient agents
MCP 三方关系用三块色区表达：Model、MCP client、MCP server，协议版本写在中间。
Beyond permission prompts: making Claude Code more secure and autonomous
最佳样本：User / Claude code / Sandbox / 外部四段泳道，权限结果用绿 Allow、红 Deny、蓝 Ask 三个 chip，端口写在连线旁。
Beyond permission prompts: making Claude Code more secure and autonomous
用三条竖栏表达信任边界，HTTP/UDP 分流到不同代理，连线即数据路径。
Building agents that reach production systems with MCP
一图讲清 Skills 与 MCP 两个扩展机制的分工：MCP 连外部，Skills 进文件系统。
Building agents: When to fold subagents into your main agent loop
“5 layers / 1 loop”把系统拆成层级，再用一个循环把 agent 行为收束。
Inference hooks: inline data loss prevention for Claude Enterprise
How an Anthropic Japan BDR runs his day with a fleet of Claude agents
How Anthropic secures its AI-native software development lifecycle
How Anthropic's sales team rebuilt inbound with Claude Managed Agents
How Claude Code works in large codebases: Best practices and where to start
How Claude Code works in large codebases: Best practices and where to start
How Curative is building a national provider network with agents on Claude
How Datadog built a “universal machine tool” for Claude Code
A guide to the anatomy of effective commerce agents
A guide to the anatomy of effective commerce agents
Using Claude Code: session management and 1M context
What is Model Context Protocol? Connect AI to your world
New in Claude Managed Agents: run agents on a schedule and store environment variables in vaults
Effective context engineering for AI agents
左右对照：prompt engineering 与 context engineering 的信息流差异，用同一套蓝/橙/绿编码。
The Claude Code guide for startups
一页拼六个 workflow pattern，每个都用同一模板：小方块 + 细箭头 + 一个菱形判断。
New in Claude Managed Agents: self-hosted sandboxes and MCP tunnels
Claude Code effort level and model selection
Claude Code effort level and model selection
Claude Code effort level and model selection
Claude Code effort level and model selection
Claude Code effort level and model selection
Claude Code effort level and model selection
Claude Code effort level and model selection
Common workflow patterns for AI agents
顺序 workflow：唯一模板，编号即顺序。
Common workflow patterns for AI agents
并行 workflow：同一模板，多条并行支路。
Common workflow patterns for AI agents
评估-优化 workflow：加入反馈回路，仍不换视觉语言。
Agent harness design: 3 patterns for harnessing Claude's intelligence
Agent harness design: 3 patterns for harnessing Claude's intelligence
Agent harness design: 3 patterns for harnessing Claude's intelligence
Agent harness design: 3 patterns for harnessing Claude's intelligence
代码与文件结构
SKILL.md、目录树、接口表、脚本打包：把“代码长什么样”画出来，而不是贴代码截图。
9 张
Equipping agents with Agent Skills
双面板：左边 agent 配置（system prompt、skills、MCP），右边虚拟机文件系统，中间用“contents live in file system”解释映射。
Equipping agents with Agent Skills
代码类图的正确做法：不截 IDE，而是把 SKILL.md 的 YAML frontmatter 与 Markdown 两段结构高亮出来。
Equipping agents with Agent Skills
Equipping agents with Agent Skills
Scaling Managed Agents with stable interfaces
极简三层：Session、Harness、Sandbox，外加 orchestration 回流。信息少但边界清楚。
Scaling Managed Agents with stable interfaces
用表格承担“接口说明”职责：component / interface / semantics 三列，比箭头图更精确。
Writing effective tools for AI agents
Steering Claude Code: when to use CLAUDE.md, skills, hooks, and subagents
把 context window 画成一张“桌面”：system prompt、skills 索引、user 消息、tool call/result 依次堆叠，解释权在两侧批注。
A guide to the anatomy of effective commerce agents
时序与生命周期
从一次请求到多阶段流水线：时间轴、阶段、门控与回退。
6 张
Introducing advanced tool use for Claude
纵向时序：传统方式 vs tool search，两条泳道并排比较 token 占用。
Introducing advanced tool use for Claude
程序化工具调用：请求/响应沿时间轴上下走，右侧标注节省的 token。
How we built Claude Code auto mode
系统架构用嵌套容器：Multi-agent 外壳里再分 orchestrator 与 subagent 池。
How we built Claude Code auto mode
两栏“全量对话 vs 分类器看到的”，把过滤/裁剪讲成信息损失。
How we built Claude Code auto mode
两阶段流水线：Agent transcript → Stage 1 classifier → Stage 2 → 输出，阶段名即动词。
How Warp builds self-improving agents on Claude
agent session 的三段状态：任务、review、后续任务，用粉色块表达“状态”而非“组件”。
