{ "schema_version": 1, "generated_at": "2026-07-06T08:02:31.448Z", "curation_scope": "Curated OpenAI and Anthropic official blogs with durable product, model, technical-practice, harness, agent workflow, eval, safety-engineering, or implementation knowledge value.", "companies": [ "anthropic", "openai" ], "topics": [ "agent", "api_reliability", "coding_agent", "context_engineering", "developer_tooling", "evals", "harness_engineering", "mcp", "responses_api", "sandbox", "structured_outputs", "tool_use", "workflow_orchestration" ], "stats": { "total_records": 6, "by_company": { "anthropic": 3, "openai": 3 }, "by_importance": { "major": 2, "foundational": 4 } }, "records": [ { "id": "openai-introducing-codex-2025-05-16", "company": "openai", "company_label": "OpenAI", "canonical_url": "https://openai.com/index/introducing-codex/", "normalized_url": "https://openai.com/index/introducing-codex", "published_at": "2025-05-16", "title_original": "Introducing Codex", "title_zh": "OpenAI 发布 Codex", "importance": "major", "content_type": "product_practice", "topics": [ "agent", "coding_agent", "developer_tooling", "sandbox" ], "summary_zh": "这篇文章是 OpenAI 把 coding agent 推向产品形态的重要节点。它的知识价值不在“又有一个代码助手”，而在 repo 级任务、隔离执行、结果交付和人类复核如何组成工程工作流。", "key_ideas": [ "coding agent 的产品边界从补全转向异步任务执行和可审查交付。", "仓库上下文、隔离运行环境和验证命令是判断可用性的关键。", "后续日报中的 Codex 更新应回到工作流、权限、评审和验证能力上理解。" ], "practice_checklist": [ "给 coding agent 任务提供明确规格、允许路径和验证命令。", "把 agent 输出纳入代码审查、测试和回滚策略，而不是直接发布。", "关注 sandbox、依赖安装、长任务恢复和 PR 交付链路。" ], "related_blog_ids": [ "openai-new-tools-building-agents-2025-03-11" ], "related_report_dates": [] }, { "id": "anthropic-claude-code-best-practices-2025-04-18", "company": "anthropic", "company_label": "Anthropic", "canonical_url": "https://www.anthropic.com/engineering/claude-code-best-practices", "normalized_url": "https://www.anthropic.com/engineering/claude-code-best-practices", "published_at": "2025-04-18", "title_original": "Claude Code: Best practices for agentic coding", "title_zh": "Claude Code：agentic coding 最佳实践", "importance": "major", "content_type": "engineering_note", "topics": [ "agent", "coding_agent", "developer_tooling", "harness_engineering" ], "summary_zh": "这篇文章的价值在于把 Claude Code 当作工程协作者而不是聊天窗口：任务拆分、上下文准备、迭代验收和工具使用方式都会影响产出质量。它适合和 Codex、Claude Code、repo 级 agent 工作流新闻互链。", "key_ideas": [ "coding agent 的效果高度依赖任务规格、仓库上下文和反馈循环。", "小步迭代、明确验证命令和人工审查比一次性大任务更可靠。", "agentic coding 的最佳实践应被固化到项目 runbook 和 harness 中。" ], "practice_checklist": [ "给 coding agent 明确目标、限制、文件边界和验收命令。", "把复杂任务拆成可独立验证的阶段。", "让 agent 在交付前运行测试并报告残余风险。" ], "related_blog_ids": [ "anthropic-building-effective-agents-2024-12-19" ], "related_report_dates": [] }, { "id": "openai-new-tools-building-agents-2025-03-11", "company": "openai", "company_label": "OpenAI", "canonical_url": "https://openai.com/index/new-tools-for-building-agents/", "normalized_url": "https://openai.com/index/new-tools-for-building-agents", "published_at": "2025-03-11", "title_original": "New tools for building agents", "title_zh": "用于构建智能体的新工具", "importance": "foundational", "content_type": "product_practice", "topics": [ "agent", "responses_api", "tool_use", "workflow_orchestration" ], "summary_zh": "这篇文章适合作为 OpenAI agent 平台化路线的基础节点：它把模型调用、工具使用、响应编排和开发者接口放进同一套构建语境，后续涉及 Responses API、Agent SDK 或工具链封装的日报都可以反向引用。", "key_ideas": [ "agent 不只是模型能力，还需要统一的工具接口和执行编排入口。", "平台原语会改变应用层对搜索、文件、计算机使用和工具调用的组织方式。", "日报中出现新的 agent 产品或 SDK 时，应判断它是能力更新还是工作流抽象更新。" ], "practice_checklist": [ "把模型响应、工具调用和状态管理作为同一条执行链设计。", "为每个工具定义权限、输入输出 schema 和失败恢复路径。", "在引入 agent SDK 前先明确哪些步骤需要模型决策，哪些步骤应保持确定性。" ], "related_blog_ids": [ "openai-structured-outputs-2024-08-06" ], "related_report_dates": [] }, { "id": "anthropic-building-effective-agents-2024-12-19", "company": "anthropic", "company_label": "Anthropic", "canonical_url": "https://www.anthropic.com/research/building-effective-agents", "normalized_url": "https://www.anthropic.com/research/building-effective-agents", "published_at": "2024-12-19", "title_original": "Building effective agents", "title_zh": "构建有效智能体", "importance": "foundational", "content_type": "best_practice", "topics": [ "agent", "evals", "harness_engineering", "workflow_orchestration" ], "summary_zh": "这篇文章适合作为 agent 工程实践的长期基准：它强调先用简单工作流，再根据任务需要增加 agent 自主性、工具和评测闭环。后续多智能体、长任务和 harness 相关新闻都可以用它校准是否真正提升工程可控性。", "key_ideas": [ "有效 agent 不等于最大自治度，很多任务应从确定性 workflow 开始。", "工具、记忆、规划和多步骤执行都需要 eval 与可观测性支撑。", "多智能体只有在分工、上下文隔离和结果合并明确时才值得引入。" ], "practice_checklist": [ "先实现可测的单路径 workflow，再增加模型决策点。", "为每个 agent 步骤记录输入、工具调用、输出和失败恢复策略。", "用任务级 eval 验证 agent 改动，而不是只看单轮回答质量。" ], "related_blog_ids": [ "anthropic-model-context-protocol-2024-11-25" ], "related_report_dates": [] }, { "id": "anthropic-model-context-protocol-2024-11-25", "company": "anthropic", "company_label": "Anthropic", "canonical_url": "https://www.anthropic.com/news/model-context-protocol", "normalized_url": "https://www.anthropic.com/news/model-context-protocol", "published_at": "2024-11-25", "title_original": "Introducing the Model Context Protocol", "title_zh": "Anthropic 发布 Model Context Protocol", "importance": "foundational", "content_type": "engineering_note", "topics": [ "agent", "context_engineering", "mcp", "tool_use" ], "summary_zh": "这篇文章是理解 MCP 生态的基础节点。它把模型上下文、外部工具和数据源连接抽象成协议问题，后续任何 MCP server、connector、agent toolchain 的新闻都可以用它解释底层动机。", "key_ideas": [ "MCP 试图把模型与工具/数据源的连接从一次性集成变成协议化接口。", "协议层价值在于复用权限、上下文注入和工具发现方式。", "判断 MCP 项目价值时，应看它解决的真实上下文来源和权限边界。" ], "practice_checklist": [ "为 MCP server 明确数据范围、认证方式和最小权限。", "把工具输出结构化，避免把长文本直接塞回上下文。", "在 agent workflow 中记录每次外部上下文注入和工具调用证据。" ], "related_blog_ids": [], "related_report_dates": [] }, { "id": "openai-structured-outputs-2024-08-06", "company": "openai", "company_label": "OpenAI", "canonical_url": "https://openai.com/index/introducing-structured-outputs-in-the-api/", "normalized_url": "https://openai.com/index/introducing-structured-outputs-in-the-api", "published_at": "2024-08-06", "title_original": "Introducing Structured Outputs in the API", "title_zh": "OpenAI API 引入结构化输出", "importance": "foundational", "content_type": "best_practice", "topics": [ "api_reliability", "structured_outputs", "tool_use" ], "summary_zh": "这篇文章的长期价值在于把“让模型按格式输出”从提示词约定推进为 API 层约束。它适合被后续日报引用到结构化抽取、工具参数生成、工作流状态写入和降低解析失败率等场景。", "key_ideas": [ "结构化输出把 JSON Schema 变成模型调用契约，而不是应用层事后修补。", "严格 schema 能降低解析、重试和下游字段缺失的工程成本。", "这类能力会影响 agent 工具调用、数据抽取和自动化工作流的可靠性基线。" ], "practice_checklist": [ "为关键模型输出定义最小但明确的 JSON Schema。", "把 schema 校验失败纳入重试、降级和人工复核路径。", "在日报生成、候选筛选和工具调用参数里优先使用结构化契约。" ], "related_blog_ids": [], "related_report_dates": [] } ] }