概述


概述是 AI Harness 专业知识库的入口章节,由六篇文档组成:介绍与定义、发展历史、架构演进、未来发展、总结、参考资料。六篇文档共同回答三个问题:AI Harness 是什么、从哪里来、现在是什么形态、将往哪里去。

介绍与定义界定 AI Harness 的定义、词源、六层能力模型与概念边界,是全库定义与口径的源头;发展历史梳理三代架构演进的时间轴与里程碑;架构演进详解最新(第三代)架构分层与 Harness Loop 主循环工程结构;未来发展前瞻技术、标准化与产业趋势(含中国 GB/Z 185 参考);总结给出阶段判断与角色建议;参考资料收录 74 条分级资料。


1. 章节导读

图 1-1|Harness 六层能力模型与 Harness Loop 单圈映射

Harness 六层能力模型与 Harness Loop 单圈映射 第三代架构分层 · 各层对应单圈六步 · 基于本文分析绘制 L1 上下文工程 单圈第 1 步 · 目标装配 L2 工具与执行 单圈第 3 步 · 工具执行 L3 编排与控制 单圈第 2 步 · 计划 / 再计划 主循环枢纽 L4 记忆与状态 单圈第 5 步 · 状态更新 L5 评估与观测 单圈第 4 步 · 观测校验 L6 治理与安全 横切关注点 · 权限与治理 结构解读:单圈六步 = L1 装配 → L3 计划 → L2 执行 → L5 校验 → L4 更新 → 回到 L3 再计划;L6 横切贯穿各层。

数据来源:基于本文分析绘制的示意图。

1.1 是什么:介绍与定义

介绍与定义是整套知识体系的总纲,界定 AI Harness 不是 Prompt、不是 Agent,而是把模型、工具、上下文、权限、记忆、评测、可观测性与人在回路组装成可控、可复现、可治理运行环境的工程化承载层。文档给出 Harness 六层能力模型(L1 上下文工程 / L2 工具与执行 / L3 编排与控制 / L4 记忆与状态 / L5 评估与观测 / L6 治理与安全),以及三代架构演进、统一术语与概念边界。

1.2 从哪里来:发展历史

发展历史以时间轴呈现 AI Harness 从提示词工程到运行时工程的三代演进,标注关键里程碑与分水岭事件,说明「竞争重心从模型层转向 Harness 层」的成因。

1.3 现在什么形态:架构演进

架构演进详解最新(第三代)架构的分层结构与横切关注点,以及 Harness Loop——基于 Harness 的智能体主循环工程结构:单圈六步(L1 目标装配 → L3 计划 → L2 工具执行 → L5 观测校验 → L4 状态更新 → 回到 L3 再计划),配跨圈五机制与终止条件五级。

1.4 往哪里去:未来发展

未来发展前瞻技术、标准化与产业三条趋势线,包含中国国内标准化专章(GB/Z 185 参考),给出四类市场角色的建议。

1.5 收束:总结

总结给出对 AI Harness 演进阶段的整体判断,以及面向不同角色(决策者、架构师、工程师、合规、市场)的行动建议。

阅读建议:先读介绍建立全局图景,再读发展历史与架构演进理解形态,最后读未来发展、总结,把握方向与行动。

2. 文档导航

概述章节六篇文档的定位与入口如下,点击链接即可进入对应文档:

文档名一句话定位链接
介绍与定义(总纲) 定义 AI Harness,给出六层能力模型、三代演进与概念边界。 介绍与定义
发展历史 三代架构演进时间轴与里程碑。 发展历史
架构演进 最新架构分层详解 + Harness Loop 主循环工程结构。 架构演进
未来发展 技术 / 标准化 / 产业趋势(含中国标准化参考)。 未来发展
总结 阶段判断与角色建议。 总结
参考资料 74 条分级参考资料。 参考资料

Overview


The Overview section is the entry point to the AI Harness knowledge base. It is composed of six documents: Intro & Definition, History, Architecture Evolution, Future Outlook, Summary, and References. Together they answer four questions: what AI Harness is, where it came from, what form it takes now, and where it is heading.

Intro & Definition establishes the definition, etymology, the six-layer capability model, and conceptual boundaries — the source of all definitions and terminology in the library. History traces the three-generation architecture evolution timeline and milestones. Architecture Evolution details the latest (third-generation) layered architecture and the Harness Loop main-loop engineering structure. Future Outlook looks ahead at technology, standardization and industrial trends (including a China GB/Z 185 reference chapter). Summary gives a stage judgement and role recommendations. References collects 74 graded sources.


1. Reading Guide

图 1-1|Harness 六层能力模型与 Harness Loop 单圈映射

Harness 六层能力模型与 Harness Loop 单圈映射 第三代架构分层 · 各层对应单圈六步 · 基于本文分析绘制 L1 上下文工程 单圈第 1 步 · 目标装配 L2 工具与执行 单圈第 3 步 · 工具执行 L3 编排与控制 单圈第 2 步 · 计划 / 再计划 主循环枢纽 L4 记忆与状态 单圈第 5 步 · 状态更新 L5 评估与观测 单圈第 4 步 · 观测校验 L6 治理与安全 横切关注点 · 权限与治理 结构解读:单圈六步 = L1 装配 → L3 计划 → L2 执行 → L5 校验 → L4 更新 → 回到 L3 再计划;L6 横切贯穿各层。

数据来源:基于本文分析绘制的示意图。

1.1 What: Intro & Definition

Intro & Definition is the master outline of the entire knowledge system. It defines AI Harness as neither a Prompt nor an Agent, but the engineering carrier that assembles models, tools, context, permissions, memory, evaluation, observability, and human-in-the-loop into a controllable, reproducible, and governable runtime. The document lays out the six-layer capability model (L1 Context Engineering / L2 Tooling & Execution / L3 Orchestration & Control / L4 Memory & State / L5 Evaluation & Observability / L6 Governance & Safety), along with the three-generation evolution and conceptual boundaries.

1.2 Where It Came From: History

History presents the three-generation evolution from prompt engineering to runtime engineering on a timeline, marking key milestones and watershed events to explain why the center of gravity of competition has shifted from the model layer to the Harness layer.

1.3 Its Current Form: Architecture Evolution

Architecture Evolution details the latest (third-generation) layered architecture and cross-cutting concerns, together with the Harness Loop — the intelligent-agent main-loop engineering structure: a single loop of six steps (L1 goal assembly → L3 planning → L2 tool execution → L5 observation/verification → L4 state update → back to L3 replanning), with five cross-loop mechanisms and five-level termination conditions.

1.4 Where It Is Heading: Future Outlook

Future Outlook looks ahead along technology, standardization and industry trend lines, including a dedicated China standardization reference chapter (GB/Z 185), and gives recommendations for four market roles.

1.5 Conclusion: Summary

Summary gives an overall judgement on the evolution stage of AI Harness, plus actionable recommendations for different roles (decision-makers, architects, engineers, compliance, and market researchers).

Recommended reading order: start with Intro to build the full picture, then read History and Architecture Evolution to understand its form, and finally read Future Outlook and Summary to grasp direction and action.

2. Document Navigation

The positioning and entry points of the six documents in this section are listed below; click a link to open the corresponding document:

DocumentOne-line PositioningLink
Intro & Definition (Master Outline) Defines AI Harness, and lays out the six-layer capability model, three-generation evolution and conceptual boundaries. Intro & Definition
History Three-generation architecture evolution timeline and milestones. History
Architecture Evolution Layered architecture details + Harness Loop main-loop engineering structure. Architecture Evolution
Future Outlook Technology / standardization / industry trends (incl. China standardization reference). Future Outlook
Summary Stage judgement and role recommendations. Summary
References 74 graded references. References