Ch23 AI Agent 与工具使用 · 信息图
信息图:Agent 核心架构
flowchart LR
P[Prompt] --> M[Memory\n短期/长期]
P --> T[Tools\nMCP/Function]
P --> PL[Plan\nCoT/ToT/ReAct]
M --> A[Agent Loop]
T --> A
PL --> A
A --> O1[Observe]
O1 --> O2[Think]
O2 --> O3[Decide]
O3 --> O4[Act]
O4 -->|重复| O1
O4 --> EX[执行环境]
EX --> FB[Feedback]
FB --> M
classDef core fill:#f9f,stroke:#333,stroke-width:2px
class A core
信息图:Agent 推理范式演进
flowchart LR
CoT[CoT\n思维链\n2022] --> SC[Self-Consistency\n多投票]
CoT --> R[ReAct\nReason+Act\nYao 2022]
R --> ToT[ToT\nTree of Thoughts\nYao 2023]
ToT --> L[LATS\nMCTS+LLM]
R --> PE[Plan-Execute\n分离规划]
R --> Re[Reflexion\n反思+记忆\n2023]
LATS --> A[Agent 生产]
PE --> A
Re --> A
classDef base fill:#e1f5ff
classDef adv fill:#fff4e1
class CoT,SC base
class R,ToT,L,PE,Re adv
信息图:MCP 协议架构
flowchart TB
H[Host\nClaude/Gemini/Cursor] -->|MCP| S[MCP Server\n1000+]
S --> R[Resources\n数据]
S --> T[Tools\n函数]
S --> P[Prompts\n模板]
H <-->|JSON-RPC| S
S -->|stdio/HTTP| EXT[外部 API/DB]
classDef core fill:#f9f,stroke:#333,stroke-width:2px
class S core
信息图:Multi-Agent 框架对比
flowchart TB
subgraph AUTOGEN[AutoGen Microsoft]
A1[ConversableAgent] --> A2[GroupChat]
A2 --> A3[异步对话]
end
subgraph CREW[CrewAI]
B1[Role] --> B2[Task]
B2 --> B3[Tool]
B3 --> B4[类似公司]
end
subgraph LG[LangGraph]
C1[State] --> C2[Node]
C2 --> C3[Edge]
C3 --> C4[状态图]
end
subgraph SMOL[Smolagents HF]
D1[Code Agent] --> D2[写 Python]
D2 --> D3[执行]
end
classDef fr fill:#e1f5ff
class AUTOGEN,CREW,LG,SMOL fr
信息图:Agent 评测体系
flowchart LR
E[Agent 评测] --> E1[能力]
E --> E2[轨迹]
E --> E3[安全]
E1 --> S1[SWE-Bench\nGitHub 修复]
E1 --> S2[WebArena\n网页操作]
E1 --> S3[OSWorld\n桌面]
E1 --> S4[GAIA\n多步推理]
E2 --> T1[Trajectory 评估]
E2 --> T2[Agent-as-Judge]
E2 --> T3[Replay 回放]
E3 --> SE1[沙盒逃逸]
E3 --> SE2[权限滥用]
E3 --> SE3[幻觉风险]
classDef eval fill:#fff4e1
class E1,E2,E3 eval
信息图:Computer Use 全景
flowchart TB
CU[Computer Use] --> A[Anthropic\n2024-10]
CU --> B[OpenAI Operator v2\n2025-09]
CU --> C[Browser-Use\n开源]
CU --> D[Skyvern\n开源]
A --> A1[桌面 + 视觉 + 工具]
B --> B1[全 OS + 1M context]
C --> C1[GitHub 8 万星]
D --> D1[企业 RPA]
A1 --> E[评测]
B1 --> E
C1 --> E
D1 --> E
E --> E1["OSWorld 38%"]
E --> E2["WebArena 58%"]
E --> E3["人类 76-78%"]
E1 --> S[安全]
E2 --> S
E3 --> S
S --> S1[沙盒 + 权限]
S --> S2[5 大风险]
数据卡片
- MCP Servers: 1000+ 官方注册 (2025-Q3)
- AutoGen: 微软 2023, GitHub 3 万+ 星
- Voyager: 第一个 LLM 自主 Minecraft 钻石 (Anthropic 2023)
- OSWorld: Claude 3.5 38.1% vs 人类 76.0%
- WebArena: Claude 3.5 Sonnet 58.1% vs 人类 78.1%
- AI Scientist: Sakana AI 自动写论文