title, type, tags, sources, last_updated
| title |
type |
tags |
sources |
last_updated |
| 7-Layer Harness Stack |
concept |
| harness-engineering |
| agentic-ai |
| system-design |
|
| Your-AI-Isn-t-Stupid---It-Just-Needs-a-Better-Harness--Lychee-Technology-Engineering-Blog |
|
2026-04-20 |
Overview
7层 Harness Stack——production-grade AI Agent 系统的分层架构规范，从认知层到约束恢复层的完整 7 层结构。
Structure
| Layer |
Name |
Core Function |
| 1 |
Cognition |
受限操作边界，role file/job description |
| 2 |
Tools |
工具输出排序/去重/token 预算截断 |
| 3 |
Contracts & Interfaces |
JSON Schema 边界验证，防 Schema Drift |
| 4 |
Orchestration |
DAG/状态机约束允许的动作 |
| 5 |
Memory & State |
Working Memory + Persistent State 分层 |
| 6 |
Evaluation & Observation |
异构验证（规则/工具/LLM-as-judge） |
| 7 |
Constraints & Recovery |
幂等重试，Context Reset 机制 |
Principles
- 模型在 Harness 内部，不直接对用户或外部世界说话
- 每个边界交叉处有显式契约：严格 JSON Schema / 类型化函数签名 / 版本化 API spec
- 每一层产生可被模型以外的东西验证的输出
Related Concepts
Source
