Skip to content

docs: 收录 2026-08-27 调研批次（#71–75 五条编号 + 观察项 38 行 + 存量回扫） #25

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
deusyu merged 1 commit into from
Aug 27, 2026
Merged

docs: 收录 2026-08-27 调研批次（#71–75 五条编号 + 观察项 38 行 + 存量回扫） #25

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion AGENTS.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,7 +23,7 @@
- [x] Phase 2：形成自己的观点（thinking/，11 篇，持续中）
- [x] Phase 3：选一个小项目实践（practice/，1 个 Ralph Demo）
- [x] Phase 4：记录反馈迭代（feedback/，1 篇，持续中）
- [x] Phase 5：输出可展示的作品（works/，34 篇翻译 + 1 篇原创 + 2 篇外部中文收录）
- [x] Phase 5：输出可展示的作品（works/，40 篇翻译 + 1 篇原创 + 2 篇外部中文收录）

> 进度详情以人类向 README.md 的"学习路线"段为准；本节是给智能体的快照。

Expand Down
22 changes: 14 additions & 8 deletions README.en.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,8 +1,8 @@
[中文](README.md) | English

![License: MIT](https://img.shields.io/badge/license-MIT-blue)
![Articles](https://img.shields.io/badge/articles-74-green)
![Translations](https://img.shields.io/badge/translations-34-orange)
![Articles](https://img.shields.io/badge/articles-79-green)
![Translations](https://img.shields.io/badge/translations-40-orange)
[![Read online](https://img.shields.io/badge/read%20online-harness.dyu.sh-c2481d)](https://harness.dyu.sh)

# Harness Engineering Study Guide
Expand Down Expand Up @@ -117,10 +117,10 @@ harness-engineering/
├── thinking/ # Phase 2: Independent analysis (11 articles)
├── practice/ # Phase 3: Hands-on experiments (1 Ralph Demo)
├── feedback/ # Phase 4: Lessons learned (1 article)
├── works/ # Phase 5: Shareable outputs (34 translations + 1 original + 2 external Chinese captures)
├── works/ # Phase 5: Shareable outputs (40 translations + 1 original + 2 external Chinese captures)
├── tools/ # Tools that reduce the 6 complexity dimensions
├── prompts/ # Validated prompts collection
└── references/ # External resource index (74 articles with deep summaries)
└── references/ # External resource index (79 articles with deep summaries)
```

Each subdirectory has its own `AGENTS.md` explaining its purpose and conventions — a direct practice of the "progressive disclosure" principle from the original article.
Expand All @@ -131,15 +131,15 @@ Each subdirectory has its own `AGENTS.md` explaining its purpose and conventions
- [x] **Phase 2: Form your own opinions** — 11 independent analyses (ongoing)
- [x] **Phase 3: Pick a small project to practice** — Ralph Demo completed (321s, $0.31)
- [x] **Phase 4: Record feedback & iterations** — 1 article (ongoing)
- [x] **Phase 5: Produce shareable work** — 34 professional translations + 1 original synthesis + 2 external Chinese captures
- [x] **Phase 5: Produce shareable work** — 40 professional translations + 1 original synthesis + 2 external Chinese captures

## 📚 Research Library

74 articles across three knowledge tracks + 2 extended readings:
79 articles across three knowledge tracks + 2 extended readings:

| Track | Coverage | Perspectives |
|-------|----------|-------------|
| AI-Era Harness Engineering | 70 articles | OpenAI → Fowler → Anthropic → LangChain → Stanford → Claude Code reverse engineering & source leak → Subagent runtime → Sensors/SPDD/ADLC → Out-of-scope, safety auditing & quality postmortems → Evaluation trilogy → Dynamic workflows → Origins (Ralph / Hashimoto) & discipline synthesis → Codex harness anatomy → Loop Engineering trilogy → Self-evolving harnesses & RSI → Formal verification → Multi-agent scaling (Cursor / C compiler) → Official containment & evals methodology → Behavior maps / DSLs / local models / outer-loop accountability → industrial-scale mechanical porting (Bun) & harness-model co-evolution (HarnessX) → long-running harness foundations & eval-environment confounders (Anthropic backfill) → harness operations metrics & reward hacking (Cursor backfill) → tool schemas are not neutral → the software-factory debate (Dex Horthy / Osmani) → agent-swarm cost economics → deleting 80% of the system prompt → a code-review-sensor benchmark (ReviewBench) |
| AI-Era Harness Engineering | 75 articles | OpenAI → Fowler → Anthropic → LangChain → Stanford → Claude Code reverse engineering & source leak → Subagent runtime → Sensors/SPDD/ADLC → Out-of-scope, safety auditing & quality postmortems → Evaluation trilogy → Dynamic workflows → Origins (Ralph / Hashimoto) & discipline synthesis → Codex harness anatomy → Loop Engineering trilogy → Self-evolving harnesses & RSI → Formal verification → Multi-agent scaling (Cursor / C compiler) → Official containment & evals methodology → Behavior maps / DSLs / local models / outer-loop accountability → industrial-scale mechanical porting (Bun) & harness-model co-evolution (HarnessX) → long-running harness foundations & eval-environment confounders (Anthropic backfill) → harness operations metrics & reward hacking (Cursor backfill) → tool schemas are not neutral → the software-factory debate (Dex Horthy / Osmani) → agent-swarm cost economics → deleting 80% of the system prompt → a code-review-sensor benchmark (ReviewBench) → empirical refutation of TDD-as-process (Böckeler) → practical loop engineering (Osmani) → an org-scale adoption snapshot (Zalando) → open neutral harnesses & white-box compaction (Pi duo) → frozen-artifact cross-model transfer (StarHarness) |
| Cloud-Native Harness.io | 2 articles | CI/CD platform architecture (same name, different meaning) |
| Efficiency Paradox & Capability Evolution | 2 articles | YDD systematic teardown + METR follow-up (measurement-methodology crisis) |
| Extended Reading | 2 articles | Context Engineering, Human-Agent collaboration |
Expand All @@ -149,12 +149,18 @@ See [references/articles.md](references/articles.md) — each article includes c
## 📖 Translations

<details>
<summary><b>34 Chinese translations of key articles</b> (click to expand)</summary>
<summary><b>40 Chinese translations of key articles</b> (click to expand)</summary>

| Translation | Original Author | Source |
|-------------|----------------|--------|
| ⭐ [Eight Years of Wanting](works/maganti-eight-years-building-ai-translation.md) | Lalit Maganti | Personal blog |
| [Evaluating code review agents with ReviewBench](works/langchain-reviewbench-translation.md) | Nick Hollon | LangChain |
| [TDD inside the agent loop - theater or actual value?](works/fowler-tdd-in-agent-loop-translation.md) | Birgitta Böckeler | martinfowler.com |
| [Practical Loop Engineering](works/osmani-practical-loop-engineering-translation.md) | Addy Osmani | AddyOsmani.com |
| [Agentic Engineering at Zalando: A Snapshot](works/zalando-agentic-engineering-translation.md) | Bartosz Ocytko | Zalando Engineering |
| [What Is a Harness?](works/pi-what-is-a-harness-translation.md) | Earendil / Pi team | earendil.com |
| [How Compaction Works in Pi](works/pi-compaction-translation.md) | Earendil / Pi team | earendil.com |
| [StarHarness: Evolving Harnesses with Stratified Search](works/arxiv-starharness-translation.md) | ServiceNow / Mila et al. | arXiv |
| [The New Rules of Context Engineering for Claude 5](works/anthropic-context-engineering-claude5-translation.md) | Thariq Shihipar | Anthropic / Claude |
| [Better Models: Worse Tools](works/ronacher-better-models-worse-tools-translation.md) | Armin Ronacher | Personal blog |
| [Rewriting Bun in Rust](works/bun-in-rust-translation.md) | Jarred Sumner | Bun Blog |
Expand Down
22 changes: 14 additions & 8 deletions README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,8 +1,8 @@
中文 | [English](README.en.md)

![License: MIT](https://img.shields.io/badge/license-MIT-blue)
![Articles](https://img.shields.io/badge/articles-74-green)
![Translations](https://img.shields.io/badge/translations-34-orange)
![Articles](https://img.shields.io/badge/articles-79-green)
![Translations](https://img.shields.io/badge/translations-40-orange)
[![在线阅读](https://img.shields.io/badge/在线阅读-harness.dyu.sh-c2481d)](https://harness.dyu.sh)

# Harness Engineering 学习指南
Expand Down Expand Up @@ -116,10 +116,10 @@ harness-engineering/
├── thinking/ # Phase 2：独立思考与质疑（11 篇）
├── practice/ # Phase 3：小项目实验（1 个 Ralph Demo）
├── feedback/ # Phase 4：踩坑与迭代心得（1 篇）
├── works/ # Phase 5：可展示的作品（34 篇翻译 + 1 篇原创 + 2 篇外部中文收录）
├── works/ # Phase 5：可展示的作品（40 篇翻译 + 1 篇原创 + 2 篇外部中文收录）
├── tools/ # 工具具像化：降低 6 维复杂度的杠杆库
├── prompts/ # 验证有效的提示词积累
└── references/ # 外部资源索引（74 篇文章深度摘要）
└── references/ # 外部资源索引（79 篇文章深度摘要）
```

每个子目录都有自己的 `AGENTS.md`，说明该目录的用途和写作约定。这本身就是原文「渐进式披露」的实践。
Expand All @@ -130,15 +130,15 @@ harness-engineering/
- [x] **Phase 2：形成自己的观点** — 11 篇独立思考（持续中）
- [x] **Phase 3：选一个小项目实践** — Ralph Demo 完成（321 秒，$0.31）
- [x] **Phase 4：记录反馈迭代** — 1 篇（持续中）
- [x] **Phase 5：输出可展示的作品** — 34 篇专业翻译 + 1 篇原创综合分析 + 2 篇外部中文收录
- [x] **Phase 5：输出可展示的作品** — 40 篇专业翻译 + 1 篇原创综合分析 + 2 篇外部中文收录

## 📚 研究资料库

跨三条知识脉络 74 篇文章 + 2 篇延伸阅读：
跨三条知识脉络 79 篇文章 + 2 篇延伸阅读：

| 脉络 | 覆盖 | 核心视角 |
|------|------|---------|
| AI 时代的 Harness Engineering | 70 篇 | OpenAI → Fowler → Anthropic → LangChain → Stanford → Claude Code 逆向与源码实锤 → Subagent runtime → 传感器/SPDD/ADLC → 越界·安全审计·质量复盘 → 评测三部曲 → 动态工作流 → 起源考据（Ralph / Hashimoto）与学科汇流 → Codex harness 解剖 → Loop Engineering 三部曲 → 自演化 harness 与 RSI → 形式化验证 → 多智能体并行规模化（Cursor / C compiler）→ 遏制与评测官方方法论 → 行为地图 / DSL / 本地模型 / 外环问责 → 工业级机械移植（Bun）与 harness-模型共演化（HarnessX）→ 长时 harness 奠基与评测环境混杂（Anthropic 存量）→ harness 运维度量与奖励作弊（Cursor 存量）→ 工具 schema 不中立 → 软件工厂之争（Dex Horthy / Osmani）→ 智能体蜂群成本经济学 → 删掉 80% 系统提示词 → 代码评审传感器基准（ReviewBench） |
| AI 时代的 Harness Engineering | 75 篇 | OpenAI → Fowler → Anthropic → LangChain → Stanford → Claude Code 逆向与源码实锤 → Subagent runtime → 传感器/SPDD/ADLC → 越界·安全审计·质量复盘 → 评测三部曲 → 动态工作流 → 起源考据（Ralph / Hashimoto）与学科汇流 → Codex harness 解剖 → Loop Engineering 三部曲 → 自演化 harness 与 RSI → 形式化验证 → 多智能体并行规模化（Cursor / C compiler）→ 遏制与评测官方方法论 → 行为地图 / DSL / 本地模型 / 外环问责 → 工业级机械移植（Bun）与 harness-模型共演化（HarnessX）→ 长时 harness 奠基与评测环境混杂（Anthropic 存量）→ harness 运维度量与奖励作弊（Cursor 存量）→ 工具 schema 不中立 → 软件工厂之争（Dex Horthy / Osmani）→ 智能体蜂群成本经济学 → 删掉 80% 系统提示词 → 代码评审传感器基准（ReviewBench）→ TDD 过程实证否定与结果度量（Böckeler）→ loop 日常落地（Osmani）→ 组织级采纳快照（Zalando）→ 开放中立 harness 与 compaction 白盒（Pi 双篇）→ 冻结工件跨模型迁移（StarHarness） |
| 云原生 Harness.io | 2 篇 | CI/CD 平台架构（同名不同义的参照） |
| 效率悖论与能力进化 | 2 篇 | YDD 系统性拆解 + METR 实验后续（测量方法论危机） |
| 延伸阅读 | 2 篇 | Context Engineering、人机协作 |
Expand All @@ -148,12 +148,18 @@ harness-engineering/
## 📖 翻译作品

<details>
<summary><b>34 篇核心文章的中文翻译</b>（点击展开）</summary>
<summary><b>40 篇核心文章的中文翻译</b>（点击展开）</summary>

| 作品 | 原作者 | 来源 |
|------|--------|------|
| ⭐ [渴望了八年，用 AI 三个月造出来](works/maganti-eight-years-building-ai-translation.md) | Lalit Maganti | 个人博客 |
| [用 ReviewBench 评测代码评审智能体](works/langchain-reviewbench-translation.md) | Nick Hollon | LangChain |
| [agent loop 里的 TDD：走形式还是真价值？](works/fowler-tdd-in-agent-loop-translation.md) | Birgitta Böckeler | martinfowler.com |
| [Practical Loop Engineering（循环的日常落地）](works/osmani-practical-loop-engineering-translation.md) | Addy Osmani | AddyOsmani.com |
| [Zalando 的 Agentic 工程快照](works/zalando-agentic-engineering-translation.md) | Bartosz Ocytko | Zalando Engineering |
| [What Is a Harness?（harness 是什么）](works/pi-what-is-a-harness-translation.md) | Earendil / Pi 团队 | earendil.com |
| [Compaction 在 Pi 里如何工作](works/pi-compaction-translation.md) | Earendil / Pi 团队 | earendil.com |
| [StarHarness：分层搜索演化企业环境 harness](works/arxiv-starharness-translation.md) | ServiceNow / Mila 等 | arXiv |
| [Claude 5 世代模型的上下文工程新规则](works/anthropic-context-engineering-claude5-translation.md) | Thariq Shihipar | Anthropic / Claude |
| [更好的模型：更差的工具](works/ronacher-better-models-worse-tools-translation.md) | Armin Ronacher | 个人博客 |
| [用 Rust 重写 Bun](works/bun-in-rust-translation.md) | Jarred Sumner | Bun Blog |
Expand Down
14 changes: 10 additions & 4 deletions prompts/deep-research-tracker.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -61,11 +61,11 @@
> 它必须自包含，因为搜索器无法访问 `references/articles.md`。
>
> **维护纪律：** 当 `references/articles.md` 新增/删除条目时，**同一次提交中**必须同步更新本节。两份内容的口径（脉络划分、篇数、产品/项目清单）应保持完全一致。
> 本节最近一次同步：2026-08-05（与 `articles.md` 当前内容对齐：74 篇文章 + 1 项已跟踪产品）。
> 本节最近一次同步：2026-08-27（与 `articles.md` 当前内容对齐：79 篇文章 + 1 项已跟踪产品）。

**核心文章 74 篇，分布于三条脉络：**
**核心文章 79 篇，分布于三条脉络：**

- **脉络一 — AI 时代 Harness Engineering（70 篇）：**
- **脉络一 — AI 时代 Harness Engineering（75 篇）：**
- OpenAI "Harness engineering"（原点，2026-02-11）/ "An open-source spec for Codex orchestration: Symphony"（2026-04-27，任务跟踪器作为控制平面）
- Fowler/Böckeler "Harness engineering for coding agent users"（2026-04-02）+ 前传备忘录（2026-02-17）
- LangChain "The Anatomy of an Agent Harness"（2026-03）/ "Continual Learning for AI Agents"（2026-04-05）/ "Agent Evaluation Readiness Checklist"
Expand Down Expand Up @@ -132,6 +132,11 @@
- Rethinking Harness Evolution 论文（arXiv 2607.12227，自动 harness 演化的第一份系统性负面结果：同等预算下不稳定优于 test-time scaling、泛化有限）
- LangChain "How We Benchmark Deep Agents"（2026-07-23）+ "IssueBench"（2026-07-20）——Harbor 评测栈：Harbor-Index 82 任务 / lite 冻结子集 / capability suite；IssueBench 15 类失败分类法与 issue 集层面判分
- LangChain / Nick Hollon "Evaluating code review agents with ReviewBench"（2026-07-31，真实 PR 评审意见策展成 59 任务 / 64 基线问题；裸 harness 最强召回 ~30%，prompt-only 调优 Luna 0.13→0.32 反超 Opus/Kimi）
- Fowler / Birgitta Böckeler "TDD inside the agent loop - theater or actual value?"（2026-08-10，agent loop 内 TDD 实证否定：盲评质量无差异、mutation score 无差异、token 3–8.5 倍；别规定过程，度量结果）
- Addy Osmani "Practical Loop Engineering"（2026-08-14，loop 系列日常落地环：踩坑实录 + 80K star 仓库分诊循环 + fine print 运维细则）
- Zalando / Bartosz Ocytko "Agentic Engineering at Zalando: A Snapshot"（2026-08-14，250+ 团队非供应商组织级快照：事故分析→PR 风险分级→33% 自动放行→lead time -20~40%）
- Earendil / Pi 团队双篇 "What Is a Harness?" + "How Compaction Works in Pi"（2026-08-20 / 08-13，harness 定义+用户侧中立主张 / compaction 白盒实现，与 Codex 端点化对照）
- StarHarness 论文（arXiv 2608.24804，ServiceNow/Mila，2026-08-25，冻结 harness 工件跨 GPT/Qwen 迁移 12 行全正 +10.7~+46.3pp；跨模型可移植性首个正面证据）
- **脉络二 — 云原生 Harness.io（2 篇）：** Harness.io 官方全局架构 / Google Cloud 集成场景
- **脉络三 — 效率悖论（2 篇）：** YDD/Miss-you "效率悖论的系统性拆解"（2026-03-03）/ METR 实验后续 + 自报调查（2026-02-24 + 2026-05-11，"慢 19%"的官方后续：弱证据转向加速 + RCT 方法论危机）

Expand Down Expand Up @@ -326,7 +331,8 @@
| 2026-07-21 | （教训来源，未做系统回扫） | 事后补收 #48 / #49 / #50 / #47 |
| 2026-07-27 | ① anthropic.com/engineering 全量列表 ② cursor.com/blog 全量 slug 清单 | #58 Effective harnesses（2025-11）、#59 基础设施噪声（2026-02）、#60 持续改进 agent harness（2026-04）、#61 奖励作弊（2026-06）——四篇均为 harness 主题正中靶心却漏网数月 |
| 2026-08-03 | ① openai.com/sitemap.xml/engineering 全量 URL+lastmod ② martinfowler.com/feed.atom（GenAI 归档） | 无漏网存量：OpenAI Engineering 分类的 harness 主题件（harness-engineering / unrolling / unlocking / symphony / windows-sandbox / core-dump / websockets / how-agents-transforming-work）均已在编号正文或观察项；其余 gpt-5.6 / atlas / sora-android / tax-agents / data-agent 为产品·基建，非 harness。Fowler 侧捞出的是增量而非存量（重构经济效益 / Conductor Developer / Fragments 07-21 / Orchestrator's Tax 草稿）——已进本批观察项 |
| 2026-08-27 | ① langchain.com/blog sitemap 全量（507 slug）② claude.com/blog sitemap 全量（228 slug，去 locale 变体） | **非零星漏网而是系统性缺口**：langchain 相关约 115 篇/此前仅收 19（漏网详查 48，含 12 篇标题直接带 harness/loop 的核心文：improving-deep-agents-with-harness-engineering、better-harness、the-art-of-loop-engineering、deep-agents 开山文等）；claude.com 相关约 80/仅收 9（漏网详查 41：skills 全系列、subagent 系列、Lessons from building Claude Code 双篇等）。处置：不逐篇消化，观察项表加 2 个"信源级"合并行，完整清单留档 translate/2026-08-27/candidates.md。方法论教训：两站列表页对 curl 均不完整**必须走 sitemap.xml**；claude.com 的 lastmod 批量刷新不可作发布日期，需逐页取 JSON-LD datePublished |

> 下一轮建议轮换：langchain.com/blog（Observability & Evals 分类）+ claude.com/blog（全量 slug）。openai.com/index 用 `curl https://openai.com/sitemap.xml/engineering/` 拿到带 lastmod 的全量 URL 最省事（2026-08-03 验证有效，比抓列表页首屏可靠）。
> 下一轮建议轮换：addyosmani.com/blog + simonwillison.net 月归档（个人博客侧尚未系统回扫过）；cursor.com/blog 可顺带复扫（本批已全量取 sitemap 115 slug 逐页核过日期，短期内干净）。openai.com/index 用 `curl https://openai.com/sitemap.xml/engineering/` 拿到带 lastmod 的全量 URL 最省事（2026-08-03 验证有效，比抓列表页首屏可靠）。
> 回扫技巧（2026-07-27 验证有效）：列表页只渲染最近若干条时，直接抓 `curl <blog>|grep -oE '/blog/[a-z0-9-]+'|sort -u` 拿全量 slug，
> 再逐个取 `datePublished` 与标题——Cursor 那四条里有两条（`continually-improving-agent-harness`、`reward-hacking-coding-benchmarks`）就是这样发现的，它们不在列表页首屏。
Loading
Loading