🔗 在线主页 / Live site  ·  📖 阅读地图 / Reading map  ·  ⚙️ HarWork demo

Building an Agent Harness

An 18-article bilingual deep dive on the engineering behind a production-grade AI Agent Harness, drawn from HarWork — a solo-built platform shipped in 49 days.

18 篇双语工程拆解 · 从 HarWork（一人 49 天打造的企业级 Agent Harness）的真实实现讲起。

Status

✅ Complete · 18/18 articles published (ZH + EN), 54 original SVG diagrams.

📖 Start here: docs/blog/reading-map.md — full bilingual reading map with topical reverse-index.

At a glance

• 287 commits / 49 days / 60,444 LOC / 110 tests — the real HarWork numbers behind the series
• 18 parts organized into 7 themes (Thesis → Core Loop → Tools → Sandbox → Storage → Streaming → Design → DevOps → Retro)
• 54 SVG diagrams, all with <title> / <desc> for accessibility
• ~180K Chinese characters + ~40K English words

Repo layout

Path	Purpose
docs/blog/reading-map.md	Entry point — bilingual reading map + reverse index
docs/blog/zh/	中文版 18 篇
docs/blog/en/	English version, 18 articles
docs/blog/assets/img/	54 original SVG diagrams
docs/blog/seo-matrix.md	Keyword × article map
docs/blog/publishing-checklist.md	Per-article release checklist
docs/blog/DECISIONS.md	Hosting / structure ADRs
scripts/blog/	Content production tooling

Reading paths

• From scratch (18 articles): 01 → 02 → ... → 18
• Skim path (8 articles): 01 → 03 → 04 → 07 → 08 → 10 → 13 → 18
• Hacker News path (4 articles): 01 → 11 → 17 → 18
• "I want to learn X" reverse index: see reading-map.md

Series themes

1. Thesis (01-02) — what an Agent Harness is, full architecture map
2. Core loop (03-06) — async generators, context compression, tool orchestration, long-term memory
3. Tools & sandbox (07-09) — protocol, permissions, hooks
4. Session storage (10-11) — 30-table schema, persistent per-user Docker
5. Streaming (12-13) — WebSocket 30s grace, multi-model routing
6. Design collaboration (14-16) — AI artifacts, variants, optimistic-lock collab
7. DevOps & retro (17-18) — canary CI/CD, 49-day solo retrospective

License

Dual-licensed — see LICENSE for the full text.

• Content (docs/, README): CC BY-SA 4.0 — free to share & adapt with attribution; derivatives must use the same license.
• Tooling (scripts/, .github/): MIT — use freely with attribution.

Disclaimer · 版权与原创声明

This series analyzes the publicly observable behavior and architecture of AI coding tools (including Claude Code) for educational purposes. All code shown is HarWork's own independent implementation — a clean-room reconstruction based on public information and observable behavior. It does not reproduce, redistribute, or quote Anthropic's source code. Claude Code, Anthropic, and other names are property of their respective owners; the engineering ideas, interfaces, and protocols discussed are not copyrightable subject matter (17 U.S.C. §102(b)).

本系列出于学习目的,分析 AI 编程工具(含 Claude Code)的可观测行为与架构设计。文中所有代码均为 HarWork 自身的独立实现——基于公开信息与可观测行为的 clean-room(净室)重建,不复制、不转载、不摘录 Anthropic 的源代码。Claude Code、Anthropic 及其它名称归各自所有者所有;所讨论的工程思想、接口与协议不属于版权保护对象。

---

🔗 中文读者：阅读地图 · seo-matrix.md 含全部中英关键词映射。