🧬 进化四阶段
LEVEL 1
Prompt Engineering
优化单次交互的指令质量
CoT · ToT · System Prompt
▼
LEVEL 2
Context Engineering
管理 Agent 能看到的信息
Progressive Disclosure · Compaction
Context Window Mgmt
Context Window Mgmt
▼
LEVEL 3
Harness Engineering
搭建完整基础设施底盘
Tools · Sandbox · Memory · MCP
Guardrails · LoopAgent
Guardrails · LoopAgent
▼
LEVEL 4
Loop Engineering
设计自动化系统取代人工
Automations · Worktrees · Skills
Sub-agents · Completion Criteria
Sub-agents · Completion Criteria
🏗️ Agent 架构五层模型
🔮 循环层 Loop Layer
决策引擎
ReAct
CoT / ToT
Reflection
Plan-Execute
LoopAgent
Loop Engineering
Completion Criteria
Sub-agent Split
↓ 循环层调用工具，依赖基础设施层提供的执行环境 ↓
⚙️ 基础设施层 Harness Layer
运行底盘
Tool Calling
Function Calling
Sandbox
MCP
Structured Output
Context Mgmt
Model Routing
Worktrees
Automations
↓ Harness 注入知识增强层的检索结果和记忆 ↓
🧠 知识增强层 Knowledge Layer
外挂大脑
RAG
Vector DB
Embedding
Chunking
Memory System
Skills (SKILL.md)
External APIs
↓ 贯穿所有层次的横切关注点 ↓
🛡️ 质量保证层 Quality Layer
横切面
Agent Evaluation
HITL
Prompt Injection 防御
Observability
Token 成本监控
Guardrails
Error Recovery
🚀 前沿扩展层 Frontier Layer
2026+
A2A Protocol
Computer Use
Multi-Model Fusion
Visual Agency
Distributed Agents