# Daily i+1 English Reading - 2026-05-27 # Daily i+1 Reading Recommendations ## Context used - 读取了你昨日(2026-05-26)的本地记录:`.../odyssey/0 收集箱/research/推文总结_2026-05-26.md`,主线是 **AI coding / computer-use agent 产品化**、**agent 工程从 demo 转向 harness/eval/部署/权限**、以及 **成本/可靠性/长任务能力**。 - 读取了你昨日(2026-05-26)的复盘:`.../odyssey/0 收集箱/Agent健身房/2026-05-26/早间复盘.md`,关注点集中在 **策略/枚举注入**、**完成态 watchdog**、**clone preflight**、**产物索引**(都属于“可审计、可回归、可规模化”的工程化摩擦)。 - 读取了你昨日(2026-05-26)的 i+1 清单:`.../odyssey/0 收集箱/每日英语i+1阅读/2026-05-26 每日英语i+1阅读.md`,为避免重复,今天不再推荐 NSA MCP 安全 PDF / prompt injection 论文 / RAG eval 入门 / web.dev long tasks。 - 未能使用:可直接查询的浏览器历史/可读导出数据源(本次没发现现成入口)。 ## Recommendations 1) Why Policy in Amazon Bedrock AgentCore chose Cedar for securing agentic workflows 2. Link: https://aws.amazon.com/blogs/security/why-policy-in-amazon-bedrock-agentcore-chose-cedar-for-securing-agentic-workflows/ 3. Topic: 给 agent 的 tool-use 加“可审计、确定性”的授权层(policy engine / schema / formal analysis) 4. Why it matches the user: 直接对齐你昨天复盘里的“RBAC/审批疲劳/策略集中化/可回归验证”,而且把“LLM 视为不可信 actor”的边界讲得很工程化 5. Why it is i+1: 句式偏安全架构与合规表达,但仍是 blog 体;你可以把每段改写成自己系统里的控制点英文描述 6. Estimated new concepts/words/chunks count: 8 7. Likely new concepts or word chunks: - defense in depth - treat the LLM as an untrusted actor - safety envelope - deterministic enforcement layer - policy authoring / policy set - formal analysis / automated reasoning - partial evaluation - approval fatigue 8. Suggested reading method: 只读“为什么要在 orchestrator 边界做授权 + Cedar analyzability”两段;每段产出 1 句英文:`We enforce X at the tool boundary to prevent Y.` 2) Mastering Agentic Techniques: AI Agent Evaluation 2. Link: https://developer.nvidia.com/blog/mastering-agentic-techniques-ai-agent-evaluation/ 3. Topic: 把 agent 评测从“感觉好用”变成“可量化指标 + 可复现任务轨迹”(成功率、工具调用、轨迹效率等) 4. Why it matches the user: 你昨天明确把叙事转到 harness/eval/部署;这篇适合当你“评测面板英文底稿” 5. Why it is i+1: i+1 主要在“评测动词搭配 + 指标定义句式”,对 TOEFL 90 友好 6. Estimated new concepts/words/chunks count: 7 7. Likely new concepts or word chunks: - evaluation-driven development - task success rate - tool-call accuracy - trajectory / rollout - failure taxonomy - regression suite - ablation (study) 8. Suggested reading method: 只抓“指标定义句”并仿写成你自己的:把 `推文日报 watchdog`、`Meegle preflight` 各写成 3 个可度量指标(英文)。 3) Agents SDK (OpenAI API docs): focus on “Sandbox agents” + “Evaluate agent workflows” 2. Link: https://developers.openai.com/api/docs/guides/agents ([platform.openai.com](https://platform.openai.com/docs/guides/agents-sdk/)) 3. Topic: agent 运行、sandbox、结果/状态、以及如何把 workflow 做成可评测/可追踪的工程单元 ([platform.openai.com](https://platform.openai.com/docs/guides/agents-sdk/)) 4. Why it matches the user: 你正在搭建“本地 workflow + 浏览器扩展 + Anki”的长期系统;这份文档更像“可落地的组件清单” 5. Why it is i+1: 文档长但结构清晰,适合按目录挑读;i+1 主要来自术语与操作动词(orchestration / guardrails / observability) 6. Estimated new concepts/words/chunks count: 6 7. Likely new concepts or word chunks: - orchestration - guardrails - observability / tracing - sandboxed execution - state management - evaluation harness 8. Suggested reading method: 只读目录级别并选 2 个小节深读:`Sandbox agents` + `Evaluate agent workflows`;每个小节输出 1 张两列表(Term → 你系统里的对应物/文件/日志字段)。 4) llm-wiki (Andrej Karpathy) 2. Link: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f 3. Topic: 把“临时对话输出”升级为“可持续演化的 Markdown wiki/知识资产”,强调编译、结构、增量与可维护性 4. Why it matches the user: 你正在把很多工作沉淀到 Odyssey/技能/产物索引里;这篇能直接给你“知识库=可编译 artifact”的英文表达 5. Why it is i+1: 观点型短段落,读起来轻;i+1 来自写作性抽象词与比喻(compile/curate/accumulate) 6. Estimated new concepts/words/chunks count: 7 7. Likely new concepts or word chunks: - incrementally compile - curated source material - durable artifact (vs. transient chat) - provenance / traceability - contradiction / unresolved links - schema-driven - review queue 8. Suggested reading method: 目标不是全读:只读“核心想法 + 工作流描述”;把它改写成你自己的 6 句英文 SOP(输入→编译→审阅→发布→回滚→索引)。 ## Vocabulary budget - Estimated daily new-item total: 8 + 7 + 6 + 7 = 28(≥20) - Back-calculate: `14678 / 28 ≈ 524` 天,约 `524 / 365 ≈ 1.44` 年 - 说明:这是“规划预算”,不是承诺;只有高复用、能写出你自己的语境例句、能迁移到 PRD/审计/复盘里的项才值得做卡。 ## How to use with Anki - 加到「英语概念卡」:优先收“你能立刻复用的句式/控制点表达”,例如 `deterministic enforcement layer`、`evaluation-driven development`、`regression suite`、`durable artifact`,每张卡必须绑定你自己的例句(来自:watchdog / preflight / artifact indexer / Odyssey 工作流)。 - 不要加:一次性专有名词堆叠、你不会在写作/文档里复用的公司/产品细节、以及你已 mastered/已 suspended 的概念。 - 「阅读词汇量」是 backlog/参考词汇库;真正需要“带语境能复述、能落到你的系统字段/流程里”的,才进入「英语概念卡」。