Skip to content

docs: 收录 2026-07 下旬调研批次(9 篇新条目 + 7 行观察项 + 3 篇翻译 + 存量回扫流程) - #16

Merged
cursor[bot] merged 1 commit into
mainfrom
cursor/curate-2026-07-scale-batch-1a25
Jul 22, 2026
Merged

docs: 收录 2026-07 下旬调研批次(9 篇新条目 + 7 行观察项 + 3 篇翻译 + 存量回扫流程)#16
cursor[bot] merged 1 commit into
mainfrom
cursor/curate-2026-07-scale-batch-1a25

Conversation

@deusyu

@deusyu deusyu commented Jul 21, 2026

Copy link
Copy Markdown
Owner

概要

curate-research 流水线收录 2026-07 下旬批次(人类闸门已确认)。编号正文 50 → 59 篇,翻译 27 → 30 篇,本批全程执行审计后新规(sourceFigureCount 声明、插图入库 works/imgs/、超链接保留、关键数字回对原文),C1–C12 全绿

新编号条目(脉络一 #47–55)

存量补课四件套(时间窗扫描的漏网,1–5 月官方重磅):

# 条目 定位
47 Anthropic《Demystifying evals for AI agents》(01-09) 评测方法论官方补件:outcome≠transcript、pass@k/pass^k、CORE-Bench 42%→95% 基准病理
48 Cursor / Wilson Lin《Scaling long-running autonomous coding》(01-14) 数百并发智能体 × 数周:扁平自协调失败史 → planner/worker/judge;FastRender 百万行浏览器
49 Anthropic / Carlini《Building a C compiler with parallel Claudes》(02-05) 16 agent 无编排者、$20k/两周成本实账、GCC oracle 拆单体任务、"为 Claude 写测试"
50 Anthropic《How we contain Claude across products》(05-25) 三隔离模式 × 五起漏掉的风险;"白名单是能力授予";先环境层遏制再模型层引导

7 月新增五条:#51 LangChain 动态子智能体+RLM(跨厂商收敛实锤)、#52 Böckeler 本地模型双备忘录、#53 Harness Handbook 论文(行为定位瓶颈,win rate +10.0/+18.9pp)、#54 Unmesh Joshi《DSLs Enable Reliable Use of LLMs》(语言层 harness)、#55 Osmani《Own the Outer Loop》(外环问责)。

新翻译(works/ +3,全部按 C10 新规)

  • works/anthropic-c-compiler-translation.md(sourceFigureCount: 0,代码块与正文链接保留)
  • works/cursor-scaling-agents-translation.md(1 图入库 works/imgs/cursor-scaling-agents/
  • works/anthropic-how-we-contain-translation.md(5 图入库 works/imgs/anthropic-how-we-contain/,逐张中文图注,正文超链接全量保留)

观察项与流程

  • 观察项 +7 行:ChatGPT Work + Codex 合并(07-09)、LangChain Prompt Caching(49–80% 节省)、Osmani《Don't Outsource the Learning》(技能形成 RCT 50% vs 67%)、The Archaeologist's Copilot、Willison llm-coding-agent、donggeking 中文教材(🔵 待实测)
  • 更新 2 行(Fragments 07-13 补 Kief Morris/Sam Ruby、Litt 演讲视频上线);中文转译段 +1(Jinyan Su 双语综述);待补充更新马东锡系列新帖线索
  • prompts/deep-research-tracker.md 新增**「核心信源存量回扫」**步骤——本批 4 篇补课暴露"只扫时间窗"的盲区,把教训固化进流程

机械化同步

  • 脉络二/三编号顺移 47–50 → 56–59(锚点 + 7 处 YDD 交叉引用同步)
  • 下游计数缓存全同步(README ×2、根/references/works AGENTS.md、deep-research-tracker)
  • bash scripts/check-consistency.sh:C1–C12 全部通过(C10 图片保真 30 篇全过、C11 表格列数、C12 作者/日期字段完备)
  • 关键数字对照原文复核:Handbook win rate 38.3%/28.3% 与 45.6%/26.7%、CORE-Bench 修复含"换更少约束的 scaffold"、钓鱼 25 试 24 中、审批率 93%、弹窗 −84% 等
Open in Web Open in Cursor 

- 编号正文 +9(50→59):存量补课四件套——Anthropic 评测方法论(#47)、
  Cursor 数百智能体规模化(#48)、Carlini 并行 Claude 造 C 编译器(#49)、
  Anthropic 遏制架构(#50);7 月新增——LangChain 动态子智能体+RLM(#51)、
  Böckeler 本地模型双备忘录(#52)、Harness Handbook 论文(#53)、
  Unmesh Joshi DSL 即 harness(#54)、Osmani 外环问责(#55)
- works/ 翻译 +3(27→30):C compiler(图 0)、Cursor scaling(图 1 入库)、
  How we contain(图 5 入库,works/imgs/ 本地嵌入 + 中文图注 + 原文超链接保留)
- 观察项 +7 行(ChatGPT Work 合并、LangChain prompt caching、Osmani 学习外包、
  Archaeologist's Copilot、llm-coding-agent、donggeking 教材 🔵)+ 2 行更新
  (Fragments 07-13 补 Kief/Sam Ruby、Litt 视频上线);中文转译段 +1(Jinyan Su)
- prompts/deep-research-tracker.md 新增「核心信源存量回扫」步骤
  (本批 4 篇补课即时间窗盲区的教训固化)
- 脉络二/三编号顺移(47-50 → 56-59,YDD 交叉引用同步),关键数字对照原文
  复核(Handbook win rate 10.0/18.9pp、CORE-Bench 42→95 等),C1–C12 全绿

Co-authored-by: Rainman <deusyu@users.noreply.github.com>
@cursor
cursor Bot marked this pull request as ready for review July 22, 2026 07:48
@cursor
cursor Bot merged commit e53dec5 into main Jul 22, 2026
1 check passed
SimonChung pushed a commit to SimonChung/harness-engineering that referenced this pull request Jul 28, 2026
5 张力对照 Böckeler 框架与 claude-code-harness v4.2 "Hokage" 实现:
Advisor 装不进 2×2、guardrail 在 hook 层融合前馈反馈、行为 Harness 是
结构性绕过、追上游占 95% 维护成本、harness 必带价值观锁定。

所有上游引用 pin 到 v4.2.0 tag 以保证可复现性 (上游已迭代到 v4.3.x)。

同步更新:
- thinking/AGENTS.md 索引追加新条目
- references/articles.md 新增条目 deusyu#16, 标注"被引版本 tag"
- prompts/deep-research-tracker.md 已知内容基线刷新到 18 篇文章 +
  已收录开源产品, Prompt B 的"8 篇文章体系"同步对齐
SimonChung pushed a commit to SimonChung/harness-engineering that referenced this pull request Jul 28, 2026
Resolves 6 drift issues found in REVIEW.md:

- articles.md: add authority header + machine-checkable counting rule
  (### N. global numbering = total article count); demote Chachamaru127
  v4.2 from deusyu#16 to a separate "已跟踪产品 / 项目" section; remove the
  "Medium 未找到" placeholder so numbered headings stay deterministic
  (now 18 contiguous entries)
- references/AGENTS.md: sync overview tables to new 16-18 numbering;
  add 脉络三 row and the product/project section
- prompts/deep-research-tracker.md: keep inline list as the de-dup
  authority for external search tools (which cannot read repo files);
  add explicit sync discipline note
- prompts/AGENTS.md: extend contract to accept both single-prompt and
  prompt-workflow shapes; add 下一步 section
- concepts/AGENTS.md: relax structure spec from rigid three-section
  template to layered guidance (single-source vs multi-source) with
  two hard rules
- README.md / README.en.md: thinking/ count 5 → 6 篇
- AGENTS.md (root): align Phase status to README (all ✅) with snapshot
  note pointing to README as source of truth
- works/AGENTS.md: add 下一步 section to fulfill root navigation
  promise
SimonChung pushed a commit to SimonChung/harness-engineering that referenced this pull request Jul 28, 2026
…e linking

- references/articles.md: insert Symphony as deusyu#16 in track 1 (renumber deusyu#16-18 to
  deusyu#17-19), sync the 4 downstream count caches (READMEs, deep-research-tracker,
  references/AGENTS.md) and per-track tables
- concepts/07-spec-as-product.md: new concept derived from Symphony's
  SPEC.md + WORKFLOW.md + multi-language-validation pattern; positioned as the
  outward-facing extension of "agent readability" (concept 04)
- Cross-source supplements weave Symphony into concepts 01 (tracker as record
  system), 02 (goals vs state-machine boundary), 05 (ticket-level throughput)
- thinking/cross-article-insights.md: case补充 to insight 1 (harness lifecycle
  with V1->V2->SPEC dispersal), 实证补充 to insight 5 (human-role expansion to
  PM/designer), 反例补充 to insight 7 (Elixir choice as evidence for tech-stack
  divergence rather than convergence)
oneRede pushed a commit to oneRede/harness-engineering that referenced this pull request Jul 28, 2026
…安全 + 大规模代码迁移 + 可治理性理论 + Zhukov Context Engineering)

新增翻译(works/):
- Anthropic Claude 5 代模型的上下文工程新规则(#62,unhobbling Claude + 删除 80% 系统提示)
- Anthropic AI 原生 SDLC 安全(#63,80% 代码 AI 编写 + 8 倍速度 + 将智能体视为内部威胁)
- Anthropic 大规模代码迁移(#64,Bun 100 万行 Zig→Rust + 修流程不修代码)
- arXiv 廉价代码昂贵判断(#65,可治理智能体软件工程案例研究)
- Zhukov Context Engineering(从 Context Window 到 Context Engineering 完整理论)

元数据同步更新:
- README.md / README.en.md:66 篇文章 + 36 篇翻译
- references/articles.md:新增 #62-65 条目,更新脉络一为 62 篇
- references/AGENTS.md:同步文章计数
- prompts/deep-research-tracker.md:同步去重清单至 2026-07-27

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

本批次延续 2026-07 下旬调研(deusyu#16),聚焦 Anthropic 官方三篇重磅:Context Engineering 范式转变、AI 原生 SDLC 的安全挑战、工业级代码迁移 playbook,及理论补充(可治理性 + 上下文工程基础理论)。
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants