参考资料

1. 使用说明

图 1-1｜参考资料体系：七大来源类 × A/B/C 可信度分级

参考资料体系：七大来源类 × A/B/C 可信度分级 分级决定引用规则 · 未确认条目集中第 9 章 · 素材截至 2026-09-12 参考资料库（W01–W102，共 102 条） 官方工程博客 W01–W20 · 20 条 厂商工程实践笔记 标准与法规 W21–W51 · 31 条 国标 / 法规 / 国际框架 学术论文 W52–W57 · 6 条 ReAct · Toolformer 等 行业报告与调研 W58–W80 · 23 条 开发者 / AI 采用率调研 开源项目 W81–W88 · 8 条 MCP / Codex / ADK 等 基准评测 W89–W93 · 5 条 SWE-bench 等评测榜 平台官方文档 W94–W102 · 9 条 平台文档 / 百科条目 未获确认清单 25 项 + 2 项整体缺口 集中汇总，待二次取证 A 级 · 官方一手来源 可直接引用，须给出处 B 级 · 权威二手来源 注明转述来源，数字二次核对 C 级 · 自述 / 社区解读 仅作线索，标 方可入正文 结构解读：102 条资料按类组织并连续编号 W01–W102，A/B/C 分级约束引用方式，未确认条目单列第 9 章待二次取证。

数据来源：基于本文分析绘制的示意图。

1.1. 可信度分级标准

本文件收录的全部资料按三级可信度分级。该分级继承自本工程检索报告的取证分级体系（A / B / C），供读者判断引用方式。

等级含义使用规则
A监管机构、标准组织、企业官方一手来源（gov.cn、nfra.gov.cn、cicpa.org.cn、tc260.org.cn、nist.gov、iso.org、darpa.mil、anthropic.com、openai.com、aaif.io、人民网、中国日报、arXiv 原文、官方开源仓库等）可直接引用，须给出出处
B权威二手来源：主流媒体与行业媒体转述、百科与综述、汇总站转引需注明“据 XX 报道 / 转述”；具体数字建议二次核对
C厂商自述、聚合站、社区与自媒体解读仅作线索；具体数字一律标 后方可入正文

1.2. 编号与引用规则

  • 编号形如 W01 起，按分类顺序连续编号，同一条目不重复编号；
  • 标注 者为链接或日期未经一手验证，引用时需二次确认；
    • 无链接条目按“文献名 + 发布机构 + 年份”著录，不虚构 URL；
    • “适用章节”以白皮书主题域标注（如“治理与风险”“发展展望”“技术架构”“行业赋能”“市场研究”“全篇”），便于按需取用。

2. 官方工程博客

编号名称机构年份等级链接适用章节
W01Effective context engineering for AI agentsAnthropic2025Ahttps://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents技术架构 / 发展展望
W02Effective harnesses for long-running agentsAnthropic2025Ahttps://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents技术架构
W03Harness design for long-running application developmentAnthropic2026Ahttps://www.anthropic.com/engineering/harness-design-long-running-apps技术架构 / 发展展望
W04Equipping agents for the real world with Agent SkillsAnthropic2025Ahttps://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills技术架构
W05Sandboxing: a safer and more autonomous approach（权限提示减少 84%）Anthropic2025Ahttps://www.anthropic.com/engineering/claude-code-sandboxing治理与风险
W06Introducing Agent Skills（2025-12-18 转为开放标准）Anthropic2025Ahttps://www.anthropic.com/news/skills技术架构
W07Introducing the Model Context ProtocolAnthropic2024Ahttps://www.anthropic.com/news/model-context-protocol技术架构
W08Donating the MCP and establishing the AAIFAnthropic2025Ahttps://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation发展展望
W09How we built our multi-agent research systemAnthropic2025Ahttps://www.anthropic.com/engineering/multi-agent-research-system 技术架构
W10Harness engineering: leveraging Codex in an agent-first worldOpenAI2026-02-11Ahttps://openai.com/index/harness-engineering/技术架构 / 发展展望
W11Function calling and other API updatesOpenAI2023Ahttps://openai.com/blog/function-calling-and-other-API-updates发展历史相关章
W12New tools for building agents（Responses API + Agents SDK）OpenAI2025Ahttps://openai.com/blog/new-tools-for-building-agents技术架构
W13Introducing Codex（Codex CLI）OpenAI2025Ahttps://github.com/openai/codex技术架构 / 发展展望
W14OpenAI co-founds the Agentic AI FoundationOpenAI2025Ahttps://openai.com/index/agentic-ai-foundation/发展展望
W15Agent Development Kit: Making it easy to build multi-agent applicationsGoogle2025Ahttps://googledevelopers.blogspot.com/en/agent-development-kit-easy-to-build-multi-agent-applications/技术架构
W16A year of open collaboration: Celebrating the anniversary of A2AGoogle Open Source Blog2026-04-16Ahttps://opensource.googleblog.com/发展展望
W17Linux Foundation Announces the Formation of the AAIFLinux Foundation2025-12-09Ahttps://aaif.io/press/linux-foundation-announces-the-formation-of-the-agentic-ai-foundation-aaif-anchored-by-new-project-contributions-including-model-context-protocol-mcp-goose-and-agents-md/发展展望
W18Linux Foundation Launches the Agent2Agent Protocol ProjectLinux Foundation2025-06-23Ahttps://www.linuxfoundation.org/press/linux-foundation-launches-the-agent2agent-protocol-project-to-enable-secure-intelligent-communication-between-ai-agents发展展望
W19AAIF 官网与新闻AAIF2025—2026Ahttps://aaif.io/发展展望
W20Cybersecurity updates: Summer 2025（Big Sleep、Timesketch + Sec-Gemini、FACADE）Google2025Ahttps://blog.google/technology/safety-security/cybersecurity-updates-summer-2025/治理与风险

3. 标准与法规

3.1. 协议与开放标准

编号名称机构年份等级链接适用章节
W21Model Context Protocol 官方站与规范（含 2026-07-28 版）MCP / AAIF2024—2026Ahttps://modelcontextprotocol.io/技术架构 / 发展展望
W22MCP Protocol Versions（五版规范演进表）MCP Ruby SDK2026Ahttps://ruby.sdk.modelcontextprotocol.io/protocol-versions/技术架构
W23Agent2Agent (A2A) 协议（开源仓库）Linux Foundation / Google2025—2026Ahttps://github.com/a2aproject/A2A发展展望
W24AGENTS.md 官方站AAIF2025—2026Ahttps://agents.md/技术架构 / 行业赋能
W25Agent Skills Specificationagentskills.io2025—2026Ahttps://agentskills.io/specification技术架构

3.2. 中国国家标准与指导性技术文件

编号名称机构年份等级链接适用章节
W26《人工智能 智能体互联》系列国家标准（GB/Z 185.1~185.7—2026）发布报道人民网2026-07-09Ahttps://finance-app.people.cn/n1/2026/0709/c1004-40757059.html发展展望
W27《人工智能 智能体互联》系列国家标准解读中国产业经济信息网（归口单位解读）2026Ahttps://cinic.org.cn/xw/zcdt/1643418.html发展展望
W28GB/Z 185—2026 落地：长三角智能体身份码节点首批发放中国日报2026-09-04Ahttps://cn.chinadaily.com.cn/a/202609/04/WS6a9a6773e4b09a165c788098.html发展展望 / 治理与风险
W29GB/T 45654—2025《网络安全技术 生成式人工智能服务安全基本要求》专家解读全国网络安全标准化技术委员会（SAC/TC260）2025Ahttps://www.tc260.org.cn/tc260/hygd1/202403/b429d868525e48c3b7d12a0ec8f82e5e.shtml治理与风险
W30GB 45438—2025《网络安全技术 人工智能生成合成内容标识方法》（强制性国标，具体条文与文本链接未获取，标 ）市场监管总局、国家标准化管理委员会2025B无（按文献名 + 机构 + 年份著录）治理与风险

3.3. 中国法律法规与监管文本

编号名称机构年份等级链接适用章节
W31《人工智能生成合成内容标识办法》（国信办通字〔2025〕2 号）国家网信办、工信部、公安部、国家广播电视总局2025Ahttps://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm治理与风险 / 行业赋能
W32《人工智能生成合成内容标识办法》解读中国政府网 / 新华社2025Ahttps://www.gov.cn/zhengce/202503/content_7014281.htm治理与风险
W33多措并举推进标识体系建设国家互联网应急中心2025Ahttps://www.cac.gov.cn/2025-09/06/c_1758880709361356.htm治理与风险
W34《生成式人工智能服务管理暂行办法》（七部门令第 15 号）国家网信办等七部门2023Ahttps://www.cac.gov.cn/2023-07/13/c_1690898327029107.htm治理与风险
W35《中华人民共和国网络安全法》修改决定（主席令第六十一号，新增第二十条）全国人民代表大会2025Ahttp://www.npc.gov.cn/c2/c30834/202601/t20260105_450980.html治理与风险
W36网络安全法修改决定专家解读中央网信办（中国网信网转载）2026Ahttps://www.cac.gov.cn/2026-01/02/c_1769093523928606.htm治理与风险
W37《银行业保险业数字金融高质量发展实施方案》国家金融监督管理总局办公厅2025-12Ahttps://www.nfra.gov.cn/cn/view/pages/ItemDetail.html?docId=1239741治理与风险 / 行业赋能
W38中注协提示会计师事务所在 2025 年年报审计中使用人工智能技术的风险防范中国注册会计师协会2026-03-05Ahttps://cicpa.org.cn/xxfb/news/202603/t20260305_65842.html治理与风险
W39《中国注册会计师审计准则问题解答第 18 号 / 第 19 号》中国注册会计师协会2025-01-07Ahttps://cicpa.org.cn/xxfb/news/202501/t20250123_65229.html治理与风险
W40ISO 37301:2021 等同转化 GB/T 35770—2022 报道中国标准化研究院2022Ahttps://www.cnis.ac.cn/bydt/zhxw/202210/t20221020_54060.html治理与风险
W41智能制造能力成熟度评估介绍（GB/T 39116-2020、GB/T 39117-2020）中国电子技术标准化研究院—Ahttps://www.cc.cesi.cn/service/show-2478.aspx行业赋能

3.4. 国际治理框架与行业标准

编号名称机构年份等级链接适用章节
W42NIST AI Risk Management Framework (AI RMF 1.0)NIST2023Ahttps://www.nist.gov/itl/ai-risk-management-framework治理与风险
W43NIST AI 600-1 Generative Artificial Intelligence ProfileNIST2024Ahttps://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence治理与风险
W44ISO/IEC 42001:2023（Information technology — Artificial intelligence — Management system）ISO/IEC JTC 1/SC 422023Ahttps://www.iso.org/standard/42001治理与风险
W45EU AI Act（Regulation (EU) 2024/1689）Implementation TimelineEuropean Commission AI Act Service Desk2024—2026Ahttps://ai-act-service-desk.ec.europa.eu/en/ai-act/timeline/timeline-implementation-eu-ai-act治理与风险
W46SR 11-7 汇编（Navigating Artificial Intelligence in Banking）Bank Policy Institute2024Bhttps://bpi.com/wp-content/uploads/2024/04/Navigating-Artificial-Intelligence-in-Banking.pdf治理与风险
W47HKMA《Supporting Adoption of Artificial Intelligence in Fighting Financial Crime》香港金融管理局2026-06-22Ahttps://brdr.hkma.gov.hk/eng/doc-ldg/current/20260622-1-EN治理与风险
W48IAASB 全球技术质量管理圆桌会议反馈汇总IAASB2026Ahttps://www.iaasb.org/news-events/2026-02/iaasb-publishes-global-roundtable-feedback-technology-and-quality-management治理与风险
W49IIA《Global Internal Audit Standards》The Institute of Internal Auditors2024Ahttps://www.theiia.org/en/standards/documents/治理与风险
W50OWASP Top 10 for LLM Applications（2025 版）OWASP GenAI Security Project2025Ahttps://genai.owasp.org/llm-top-10/治理与风险
W51MITRE ATLASMITRE2023—2026Ahttps://atlas.mitre.org治理与风险

4. 学术论文

编号名称作者 / 机构年份等级链接适用章节
W52ReAct: Synergizing Reasoning and Acting in Language ModelsYao 等（Princeton / Google Brain）2022（ICLR 2023）Ahttps://arxiv.org/abs/2210.03629发展历史相关章
W53Toolformer: Language Models Can Teach Themselves to Use ToolsSchick 等（Meta AI）2023（NeurIPS 2023）Ahttps://arxiv.org/abs/2302.04761发展历史相关章
W54SWE-bench: Can Language Models Resolve Real-World GitHub Issues?Jimenez、Yang 等2023（ICLR 2024 Oral）Ahttps://arxiv.org/abs/2310.06770技术架构 / 基准评测
W55Gorilla: Large Language Model Connected with Massive APIsUC Berkeley2023Ahttps://arxiv.org/abs/2305.15334技术架构
W56A Survey of Context Engineering for Large Language ModelsMei 等2025Bhttps://arxiv.org/abs/2507.13334技术架构 / 发展展望
W57CyberSentinel-LLM（SOC 智能体信任度调研）Tech Science Press（CMC vol.89 no.1）2025Bhttps://www.techscience.com/cmc/v89n1/68397/html治理与风险

5. 行业报告与调研

编号名称机构年份等级链接适用章节
W582025 Stack Overflow Developer SurveyStack Overflow2025-07-29Ahttps://survey.stackoverflow.co/2025/发展展望
W59Stack Overflow 2025 Developer Survey 官方新闻稿Stack Overflow2025Ahttps://stackoverflow.co/company/press/archive/stack-overflow-2025-developer-survey/发展展望
W60DORA 2025 State of AI-assisted Software DevelopmentGoogle Cloud / DORA2025-09Ahttps://dora.dev/dora-report-2025 发展展望
W61The State of AI 2025: Agents, Innovation, and TransformationMcKinsey2025-11-05Ahttps://www.mckinsey.com/ 发展展望
W62AI Adoption Stats & Trends（DORA / McKinsey / JetBrains 汇总）daily.dev2026-07-19Bhttps://daily.dev/agentic-ai-hub/ai-adoption-stats-trends/发展展望
W63Cost of a Data Breach Report 2025IBM / Ponemon Institute2025-07-30Ahttps://newsroom.ibm.com/2025-07-30-ibm-report-13-of-organizations-reported-breaches-of-ai-models-or-applications,-97-of-which-reported-lacking-proper-ai-access-controls治理与风险
W64DARPA AI Cyber Challenge 决赛结果DARPA2025Ahttps://www.darpa.mil/news/2025/aixcc-results治理与风险 / 技术架构
W65Harness engineering for coding agent usersBirgitta Böckeler / martinfowler.com2026Ahttps://martinfowler.com/articles/harness-engineering.html技术架构 / 发展展望
W66Humans and Agents in Software Engineering Loopsmartinfowler.com2026Ahttps://martinfowler.com/articles/exploring-gen-ai/humans-and-agents.html发展展望
W67My AI Adoption JourneyMitchell Hashimoto2026-02-05Ahttps://mitchellh.com/writing/my-ai-adoption-journey技术架构
W68拥抱智能浪潮 泳向变革深处（新华社采编助手、芒果大模型、人民网语料库）新华社2025Ahttps://www.news.cn/20251114/1c01598d836c449bbfd54267b6ecea6d/c.html行业赋能
W69主流媒体所办新媒体发展研究报告（2024-2025）人民网2025Ahttps://sc.people.com.cn/BIG5/n2/2025/1030/c345167-41396739.html行业赋能
W70AI 要给微短剧“洗牌”？（中国网络视听协会《微短剧创作指引》转引）人民日报2026Ahttps://kpzg.people.com.cn/n1/2026/0511/c404214-40717117.html行业赋能
W71Premium Times adopts Google NotebookLM to streamline newsroom processesINMA2025Ahttps://www.inma.org/blogs/conference/post.cfm/premium-times-adopts-google-notebook-lm-to-streamline-newsroom-processes行业赋能
W72《技术不是侵权“挡箭牌” 法院这样认定 AI“盗脸”》（北京互联网法院 2026-03 生效判决报道）新华社《经济参考报》2026-04-17Bhttp://dz.jjckb.cn/www/pages/webpage2009/html/2026-04/17/content_115180.htm治理与风险
W73《e案e审丨短剧角色 AI 换脸“神似”知名演员》北京互联网法院供稿 / 澎湃新闻2026Ahttps://www.thepaper.cn/newsDetail_forward_32799628治理与风险
W74JPMorgan Chase 2016 年年报（COiN 合同智能平台）JPMorgan Chase2016Ahttps://reports.jpmorganchase.com/investor-relations/2016/ar-ceo-letter-matt-zames.htm行业赋能
W75A&O Shearman 获 FT Innovative Lawyers Awards 官方新闻稿A&O Shearman（Allen Overy）2024-09-12Ahttps://www.allenovery.com/en/news/ao-shearman-wins-most-innovative-law-firm-in-europe-at-the-ft-innovative-lawyers-awards行业赋能
W76法信法律基座模型官方介绍人民法院出版社2024—2025Ahttps://www.faxin.cn/html/about/about.aspx行业赋能
W77中央网信办生成式 AI 备案用户规模报道（490+ 款 / 2.3 亿用户）央视网2025-09Bhttps://big5.cctv.com/gate/big5/news.cctv.cn/2025/09/01/ARTI3ZlXK7MyM39Pm3PuZ5Hm250901.shtml行业赋能
W782025 年上市银行 AI 应用综合报道（工银智涌、邮智等）新华网2025-09-10Bhttps://www.xinhua.org/20250910/4e421fdeeb2242fc97299d73b2d62f09/c.html行业赋能
W79券商 2025 年报 IT 投入统计（34 家 / 275.9 亿元口径）证券时报2026-04Bhttps://www.stcn.com/article/detail/3731822.html行业赋能
W80券商 AI 投研与合规实践（21 世纪经济报道专题）21 世纪经济报道2026-04-27Bhttps://www.21jingji.com/article/20260427/herald/b0e4a124fc6808e55600d6fcd453da08.html行业赋能
W80aClaude Code Changelog（v2.1.257—269 官方版本记录，2026-09）Anthropic 官方2026-09-13 快照核实Ahttps://code.claude.com/docs/en/changelog发展展望
W80bOpenAI API Changelog（2026-09-10 条目：Agents API 公测；GPT-Live 1 GA；2026-09-03：GPT-6 Astra GA）OpenAI 官方2026-09-10Ahttps://developers.openai.com/api/docs/changelog（2026-09-13 快照修正：公测日期由媒体口径 09-11 修正为官方口径 09-10）产业格局 / 发展展望
W80gMCP 路线图（2026-08-22：智能体消息原语 / HTTP 原生传输统一 / 智能体身份）MCP 官方2026-08-22Ahttps://modelcontextprotocol.io/development/roadmap发展展望
W80hCursor Projects（Beta）与 Self-Hosted Machines 更新日志Cursor 官方2026-09-10 / 09-02Ahttps://cursor.com/changelog（能力描述经 Changelog 汇总站转述核对）发展展望
W80iOpenAI 研究组织智能体使用数据（3.1 智能体工作日/人工作日；中位研究员日耗超 600 美元）OpenAI2026-09-06A官方口径；原始博文 URL [待核实]发展展望
W80j超维动力（Kinetix AI）超 5 亿元天使+轮融资（祥峰资本领投）中国新闻网广东频道 / 新京报贝壳财经2026-09-11Bhttps://www.gd.chinanews.com.cn/2026/2026-09-11/449585.shtml发展展望
W80kAccomplish 披露 Claude Code macOS 沙箱逃逸细节（07-13 上报，v2.1.247 修复）Accomplish2026-09-11B安全厂商技术披露，原文 URL [待核实]治理与风险
W80lGemini 3.8 Flash 发布（HLE-Verified 54.9%、DeepSWE v1.1 约 73.8%；$0.75/$3.75 每百万 token；同步发布 Flash Cyber）Google 官方公告 / Info Data（il Sole 24 Ore）2026-09-02 / 09-12A / B官方原文 URL [待核实]；细节核对 https://www.infodata.ilsole24ore.com/2026/09/12/come-e-fatto-gemini-3-8-flash发展展望
W80mNVIDIA 协议收购 Hugging Face（约 129.3 亿美元；2026-09-02 签署、09-03 宣布、预计 2027 上半年交割）ECM Source（引 SEC 8-K）/ Inside AI2026-09-03A（SEC 文件转述）https://ecmsource.com/nvidia-hugging-face-12-9-billion-acquisition-september-2026/（260914 快照升级：两源口径冲突关闭）发展展望 / 产业格局
W80nOpenAI 智能体滥用 RubyGems 事件复盘（2026-05 发生；RubyGems 冻结注册四天）Syntackle（The Weekly Diff #10，第三方调查汇总）2026-09 上旬Bhttps://syntackle.com/blog/the-ai-slowdown-pact-nvidia-s-13b-hugging-face-deal-the-rubygems-attack-the-weekly-diff-10治理与风险
W80oGitHub Copilot 2026-09 动态（HydraFusion、Jira 集成、代码评审 ensemble：+47%/+31%/-8%；模型退役 2026-10-02）GitHub 官方 Changelog（经 AI Coding Roundup 核对）2026-09-07 / 09-11A / B（效果数字厂商自报）https://oday-bakkour.com/blog/ai-coding-roundup-september-13-2026发展展望
W80pMCP ToolAnnotations 默认值分析（未标注工具默认视为可破坏；7 个记忆服务器普查）deniz.in（第三方分析）；MCP 2026-07-28 schema 原文 A 级2026-09-12Bhttps://deniz.in/mcp-s-annotation-defaults-make-unannotated-tools-implicitly-destructive治理与风险
W80qAWS Kiro 学生计划（18 国 132 所高校免费一年、每月 1,000 credits）AWS 官方新闻2026-09 上中旬（具体日期 [待核实]）Ahttps://www.aboutamazon.com/news/aws/amazon-expands-free-kiro-access-students-universities发展展望
W80cKiro Web 可用性与计费口径（官方 FAQ 确认；GA 媒体口径 2026-09-01）AWS / Kiro 官方2026-09-14 快照核实Ahttps://kiro.dev/faq（GA 具体日期 [待核实]）发展展望
W80dFigure AI 与 Nscale 算力合作（初始承诺 35 亿美元、规划至多 10 万颗 Vera Rubin GPU，Nscale 战略投资 Figure）具身智能应用与投融资周报（2026.9.2—9.8）2026-09 上旬B行业周报佐证，官方公告未见，[待核实] 保留；原文 URL [待核实]发展展望
W80f云蝶科技 RoboForge 系统级具身大脑（Harness 模块记录任务全流程证据链；与南洋理工大学联合研发；基准登顶为厂商自报）南方+ / 云蝶科技官网2026-09-03—06B + Chttps://www.nfnews.com/content/lyKmjgJ765.html；http://www.cloudbutterfly.com.cn/newsshow/post-227.html（更名更正：原条目误记「云蝶数据」）发展展望

6. 开源项目

编号名称机构 / 作者等级链接适用章节
W81Model Context Protocol 规范与 SDK（GitHub 组织）MCP / AAIFAhttps://github.com/modelcontextprotocol技术架构
W82OpenAI Codex CLIOpenAIAhttps://github.com/openai/codex技术架构 / 发展展望
W83Claude Agent SDKAnthropicAhttps://github.com/anthropics/claude-agent-sdk-python技术架构
W84Google Agent Development Kit (ADK)GoogleAhttps://github.com/google/adk-python技术架构
W8512-Factor AgentsHumanLayer（Dex Horthy）Ahttps://github.com/humanlayer/12-factor-agents 技术架构 / 发展展望
W86Terminal-Bench / Harbor 框架Stanford / Laude Institute / 社区Ahttps://github.com/harbor-framework/terminal-bench基准评测
W87@anthropic-ai/sandbox-runtimeAnthropicAhttps://www.npmjs.com/package/@anthropic-ai/sandbox-runtime治理与风险
W88ComfyUI（开源节点式生成工作流）Comfy OrgAhttps://github.com/comfyanonymous/ComfyUI行业赋能

7. 基准评测

编号名称机构年份等级链接适用章节
W89SWE-bench 官方站与排行榜Princeton / 社区2023—2026Ahttps://www.swebench.com/技术架构 / 发展展望
W90Terminal-Bench（官方站）Stanford / Laude Institute2025—2026Ahttps://www.tbench.ai/ 技术架构 / 发展展望
W91SWE-bench Verified 进展时间线 2023—2026（含高冲击量化数字，须标 ）AgentMarketCap2026-04-09Chttps://agentmarketcap.ai/blog/2026/04/09/swe-bench-verified-progress-timeline-2023-2026技术架构
W92Terminal-Bench: The CLI Autonomy Standard（C 级，数字须标 ）AgentMarketCap2026-04-09Chttps://agentmarketcap.ai/blog/2026/04/09/terminal-bench-cli-autonomy-standard-coding-agents技术架构
W93Artificial Analysis（图像 / 视频生成模型第三方评测榜单）Artificial Analysis2025—2026Bhttps://artificialanalysis.ai/市场研究

8. 平台官方文档

编号名称机构年份等级链接适用章节
W94Claude Code 官方文档 · SandboxingAnthropic2026Ahttps://code.claude.com/docs/zh-TW/sandboxing治理与风险
W95Claude Code 官方文档 · Choose a sandbox environmentAnthropic2026Ahttps://code.claude.com/docs/en/sandbox-environments治理与风险
W96火山引擎方舟控制台 Doubao-Seedance-2.0 系列字节跳动 / 火山引擎2026Ahttps://console.volcengine.com/ark/region:ark+cn-beijing/model/detail?Id=doubao-seedance-2-0市场研究
W97即梦 AI 官网字节跳动（剪映团队）2024—2026Ahttps://jimeng.jianying.com/市场研究
W98MCP 规范演进条目（docs.rs mcpkit ProtocolVersion）mcpkit2026Ahttps://docs.rs/mcpkit/latest/enum.ProtocolVersion.html技术架构
W99Model Context Protocol（百科条目，规范演进与三角色）Wikipedia / Klu—Bhttps://en.wikipedia.org/wiki/Model_Context_Protocol ；http://klu.ai/glossary/model-context-protocol技术架构
W100Agent2Agent (A2A)（百科条目）Wikipedia—Bhttps://en.wikipedia.org/wiki/Agent2Agent发展展望
W101Harness 架构（百科条目）百度百科2026Bhttps://baike.baidu.com/item/Harness%E6%9E%B6%E6%9E%84/67704948全篇
W102Agent Harness：2026 年 AI 工程的核心范式（框架性论述可作线索，数字标 ）腾讯云开发者社区2026Chttps://developer.cloud.tencent.com/article/2698416全篇

9. 本白皮书引用但未获权威确认的条目清单

以下条目在本工程检索中未能取得 A/B 级一手来源、或存在来源间冲突。白皮书正文引用时均已就地标注 或并列呈现，此处集中汇总，供后续修订与二次取证使用。

序号条目问题描述影响章节
1GB/Z 185—2026 发布日期2026-05-22（中国日报）/ 2026-06-26（百度百科）/ 2026-07-09（人民网“近日发布”）三种口径并存，需以官方正式公告为准发展展望
2GB 45438—2025《网络安全技术 人工智能生成合成内容标识方法》标准编号与条文在早期检索中未获官方文本链接；其元数据字段与显式标识规格来自本工程市场研究组检索所得的公开转述治理与风险3《标识办法》第七条（应用程序分发平台核验义务）具体条文未直接核对官方发布文本原文治理与风险
4GB/T 45654—2025 “附录列明 31 类风险”“生成合规内容合格率不低于 90%”二级解读，未查证标准原文，本白皮书未采用治理与风险
5即梦 AI 2026-04-28 被查处的官方通报与后续整改依据公开报道与百科条目整理，未见行政处罚决定书治理与风险6北京互联网法院 2026-03 生效 AI 换脸判决的完整案号公开报道未披露治理与风险
7EU AI Act 时间线的 AI Omnibus 调整（2026-05 临时协议）中文二手来源，按“拟议”处理，以欧盟官方服务台为准治理与风险
8NIST AI 600-1 的 12 类风险完整清单另有来源称 13 类，未取得原文清单治理与风险9最高人民法院《关于规范和加强人工智能司法应用的意见》全称、发布日期与条款仅取得转述治理与风险10《金融领域科技伦理指引》文号与条款；《人工智能算法金融应用评价规范》《人工智能算法金融应用信息披露指南》的 JR/T 编号未获取治理与风险11中国医疗 AI 专项监管文本（医用软件 / 辅助诊断审评要求）未检索到条文级来源，正文已如实声明治理与风险12券商 IT 投入两套统计口径（34 家 / 275.9 亿元 / 6.2% 与 28 家 / 约 250 亿元 / 6%）口径冲突，本白皮书采用 34 家口径并注明出处行业赋能13DORA 2025、McKinsey《The State of AI 2025》、JetBrains 2025 的全部数字与官方 URL均来自汇总站转引，未验证官方报告原文发展展望14中国厂商动向（DeepSeek 2026-05-20 组建 Harness 团队、小米 2026-06-11 MiMo Code V0.1.0、灵犀智涌 2026-08 ROSS）均仅见于百科词条转述发展展望15Codex CLI Rust 重写比例（2026 年初约 95%）B 级来源发展展望16架构重构案例数字（Manus 六个月五次重构、LangChain 三次重设计、Vercel 移除 80% 工具）B/C 级来源发展展望
17上下文退化的量化数字（“性能下降超 45%”等）C 级来源，未获 A/B 级确认发展展望
18Harness 层自身的市场规模未检索到权威测算，属 [待填写]发展展望19AGNTCY 项目仅见于综述提及，暂无权威一手资料发展展望20ISO/IEC 层面的智能体互联国际标准未检索到已发布或已立项的标准编号发展展望21"Agent Harness"术语的首创者与首次出现出处未找到确切一手文献；可确认 Anthropic 2025 年已用 "harness" 描述 Claude Agent SDK、OpenAI 2026-02-11 将 "Harness engineering" 推向主流全篇22SWE-bench Verified 与 Terminal-Bench 2.0 的 2026 年榜单数字（W91、W92）C 级来源，引用必须标 技术架构23Anthropic 多智能体研究系统博客链接（W09）与 12-Factor Agents 仓库链接（W85）链接未验证可访问性，标 技术架构24中国银行业 AI 风控效果指标（银行案例量化多为券商研究转述，非年报原文逐字核对）B 级行业赋能25中国网络视听协会《微短剧创作指引》原文文号仅经人民日报转引行业赋能

信息缺口声明

除上表 25 项外，另有两点整体性缺口需向读者说明：

  1. 时间截面：本参考资料全部素材截至 2026-09-14 快照（2026-09-12/13/14 三轮增量检索已并入 W80a—W80q，信息截止 2026-09-13）。智能体领域的法规、标准与产品更新频繁，正式引用前应回查各来源的最新版本。
  2. 检索范围：本白皮书采用“基于既有资料”模式，素材来自本工程既有的调研文档与检索报告（覆盖概述、行业赋能八大组、市场研究七组的全部已生成文档），未新增联网检索；因此本清单只汇总既有检索已识别的缺口，不排除存在未识别的缺口。

References

1. Instructions

图 1-1｜参考资料体系：七大来源类 × A/B/C 可信度分级

参考资料体系：七大来源类 × A/B/C 可信度分级 分级决定引用规则 · 未确认条目集中第 9 章 · 素材截至 2026-09-12 参考资料库（W01–W102，共 102 条） 官方工程博客 W01–W20 · 20 条 厂商工程实践笔记 标准与法规 W21–W51 · 31 条 国标 / 法规 / 国际框架 学术论文 W52–W57 · 6 条 ReAct · Toolformer 等 行业报告与调研 W58–W80 · 23 条 开发者 / AI 采用率调研 开源项目 W81–W88 · 8 条 MCP / Codex / ADK 等 基准评测 W89–W93 · 5 条 SWE-bench 等评测榜 平台官方文档 W94–W102 · 9 条 平台文档 / 百科条目 未获确认清单 25 项 + 2 项整体缺口 集中汇总，待二次取证 A 级 · 官方一手来源 可直接引用，须给出处 B 级 · 权威二手来源 注明转述来源，数字二次核对 C 级 · 自述 / 社区解读 仅作线索，标 方可入正文 结构解读：102 条资料按类组织并连续编号 W01–W102，A/B/C 分级约束引用方式，未确认条目单列第 9 章待二次取证。

数据来源：基于本文分析绘制的示意图。

1.1. Credibility rating criteria

All materials included in this document are rated on a three-level credibility scale. This scale is inherited from the evidence-grading system (A / B / C) of this project's retrieval report, and is provided for readers to judge how an item may be cited.

LevelMeaningUsage rule
AOfficial first-hand sources from regulators, standards bodies, and enterprises (gov.cn, nfra.gov.cn, cicpa.org.cn, tc260.org.cn, nist.gov, iso.org, darpa.mil, anthropic.com, openai.com, aaif.io, 人民网, 中国日报, arXiv originals, official open-source repositories, etc.)May be cited directly; the source must be given
BAuthoritative second-hand sources: relays by mainstream and industry media, encyclopedias and surveys, reprints by aggregation sitesMust note “as reported / relayed by XX”; specific figures recommended for secondary verification
CVendor self-reports, aggregation sites, and community or self-media interpretationsFor leads only; specific figures must uniformly be marked [To be verified] before entering the main text

1.2. Numbering and citation rules

  • Numbers start in the form of W01, numbered consecutively by category order; the same item is never numbered twice;
  • Items marked [To be verified] have links or dates not yet first-hand verified and require second confirmation when cited;
    • Items without links are catalogued as “document title + issuing organization + year” without fabricating a URL;
    • “Applicable sections” is annotated by whitepaper topic domain (e.g., “Governance and Risk”, “Outlook”, “Technical Architecture”, “Industry Enablement”, “Market Research”, “Entire Document”), for convenient on-demand use.

2. Official Engineering Blogs

No.NameOrganizationYearLevelLinkApplicable sections
W01Effective context engineering for AI agentsAnthropic2025Ahttps://www.anthropic.com/engineering/effective-context-engineering-for-ai-agentsTechnical Architecture / Outlook
W02Effective harnesses for long-running agentsAnthropic2025Ahttps://www.anthropic.com/engineering/effective-harnesses-for-long-running-agentsTechnical Architecture
W03Harness design for long-running application developmentAnthropic2026Ahttps://www.anthropic.com/engineering/harness-design-long-running-appsTechnical Architecture / Outlook
W04Equipping agents for the real world with Agent SkillsAnthropic2025Ahttps://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skillsTechnical Architecture
W05Sandboxing: a safer and more autonomous approach (permission prompts reduced by 84%)Anthropic2025Ahttps://www.anthropic.com/engineering/claude-code-sandboxingGovernance and Risk
W06Introducing Agent Skills (became an open standard on 2025-12-18)Anthropic2025Ahttps://www.anthropic.com/news/skillsTechnical Architecture
W07Introducing the Model Context ProtocolAnthropic2024Ahttps://www.anthropic.com/news/model-context-protocolTechnical Architecture
W08Donating the MCP and establishing the AAIFAnthropic2025Ahttps://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundationOutlook
W09How we built our multi-agent research systemAnthropic2025Ahttps://www.anthropic.com/engineering/multi-agent-research-system Technical Architecture
W10Harness engineering: leveraging Codex in an agent-first worldOpenAI2026-02-11Ahttps://openai.com/index/harness-engineering/Technical Architecture / Outlook
W11Function calling and other API updatesOpenAI2023Ahttps://openai.com/blog/function-calling-and-other-API-updatesRelated sections on development history
W12New tools for building agents (Responses API + Agents SDK)OpenAI2025Ahttps://openai.com/blog/new-tools-for-building-agentsTechnical Architecture
W13Introducing Codex (Codex CLI)OpenAI2025Ahttps://github.com/openai/codexTechnical Architecture / Outlook
W14OpenAI co-founds the Agentic AI FoundationOpenAI2025Ahttps://openai.com/index/agentic-ai-foundation/Outlook
W15Agent Development Kit: Making it easy to build multi-agent applicationsGoogle2025Ahttps://googledevelopers.blogspot.com/en/agent-development-kit-easy-to-build-multi-agent-applications/Technical Architecture
W16A year of open collaboration: Celebrating the anniversary of A2AGoogle Open Source Blog2026-04-16Ahttps://opensource.googleblog.com/Outlook
W17Linux Foundation Announces the Formation of the AAIFLinux Foundation2025-12-09Ahttps://aaif.io/press/linux-foundation-announces-the-formation-of-the-agentic-ai-foundation-aaif-anchored-by-new-project-contributions-including-model-context-protocol-mcp-goose-and-agents-md/Outlook
W18Linux Foundation Launches the Agent2Agent Protocol ProjectLinux Foundation2025-06-23Ahttps://www.linuxfoundation.org/press/linux-foundation-launches-the-agent2agent-protocol-project-to-enable-secure-intelligent-communication-between-ai-agentsOutlook
W19AAIF official website and newsAAIF2025—2026Ahttps://aaif.io/Outlook
W20Cybersecurity updates: Summer 2025 (Big Sleep, Timesketch + Sec-Gemini, FACADE)Google2025Ahttps://blog.google/technology/safety-security/cybersecurity-updates-summer-2025/Governance and Risk

3. Standards and Regulations

3.1. Protocols and Open Standards

No.NameOrganizationYearLevelLinkApplicable sections
W21Model Context Protocol official site and specification (including the 2026-07-28 version)MCP / AAIF2024—2026Ahttps://modelcontextprotocol.io/Technical Architecture / Outlook
W22MCP Protocol Versions (evolution table of the five specification versions)MCP Ruby SDK2026Ahttps://ruby.sdk.modelcontextprotocol.io/protocol-versions/Technical Architecture
W23Agent2Agent (A2A) protocol (open-source repository)Linux Foundation / Google2025—2026Ahttps://github.com/a2aproject/A2AOutlook
W24AGENTS.md official siteAAIF2025—2026Ahttps://agents.md/Technical Architecture / Industry Enablement
W25Agent Skills Specificationagentskills.io2025—2026Ahttps://agentskills.io/specificationTechnical Architecture

3.2. Chinese National Standards and Guiding Technical Documents

No.NameOrganizationYearLevelLinkApplicable sections
W26Release report on the 《人工智能 智能体互联》 series of national standards (GB/Z 185.1~185.7—2026)人民网2026-07-09Ahttps://finance-app.people.cn/n1/2026/0709/c1004-40757059.htmlOutlook
W27Interpretation of the 《人工智能 智能体互联》 series of national standards中国产业经济信息网 (interpretation by the responsible body)2026Ahttps://cinic.org.cn/xw/zcdt/1643418.htmlOutlook
W28GB/Z 185—2026 takes effect: first batch of agent identity code nodes issued in the Yangtze River Delta中国日报2026-09-04Ahttps://cn.chinadaily.com.cn/a/202609/04/WS6a9a6773e4b09a165c788098.htmlOutlook / Governance and Risk
W29Expert interpretation of GB/T 45654—2025 《网络安全技术 生成式人工智能服务安全基本要求》全国网络安全标准化技术委员会 (SAC/TC260)2025Ahttps://www.tc260.org.cn/tc260/hygd1/202403/b429d868525e48c3b7d12a0ec8f82e5e.shtmlGovernance and Risk
W30GB 45438—2025 《网络安全技术 人工智能生成合成内容标识方法》 (mandatory national standard; specific clauses and text link not obtained, marked [To be verified])市场监管总局、国家标准化管理委员会2025BNone (catalogued as document title + organization + year)Governance and Risk

3.3. Chinese Laws, Regulations, and Regulatory Texts

No.NameOrganizationYearLevelLinkApplicable sections
W31《人工智能生成合成内容标识办法》 (国信办通字〔2025〕2 号)国家网信办、工信部、公安部、国家广播电视总局2025Ahttps://www.cac.gov.cn/2025-03/14/c_1743654684782215.htmGovernance and Risk / Industry Enablement
W32Interpretation of the 《人工智能生成合成内容标识办法》中国政府网 / 新华社2025Ahttps://www.gov.cn/zhengce/202503/content_7014281.htmGovernance and Risk
W33Multi-pronged efforts to advance the construction of the labeling system国家互联网应急中心2025Ahttps://www.cac.gov.cn/2025-09/06/c_1758880709361356.htmGovernance and Risk
W34《生成式人工智能服务管理暂行办法》 (Order No. 15 of seven departments)国家网信办等七部门2023Ahttps://www.cac.gov.cn/2023-07/13/c_1690898327029107.htmGovernance and Risk
W35Decision to amend the 《中华人民共和国网络安全法》 (Presidential Order No. 61, adding Article 20)全国人民代表大会2025Ahttp://www.npc.gov.cn/c2/c30834/202601/t20260105_450980.htmlGovernance and Risk
W36Expert interpretation of the Cybersecurity Law amendment decision中央网信办 (reprinted by 中国网信网)2026Ahttps://www.cac.gov.cn/2026-01/02/c_1769093523928606.htmGovernance and Risk
W37《银行业保险业数字金融高质量发展实施方案》国家金融监督管理总局办公厅2025-12Ahttps://www.nfra.gov.cn/cn/view/pages/ItemDetail.html?docId=1239741Governance and Risk / Industry Enablement
W38中注协 reminder on risk prevention when accounting firms use AI technology in 2025 annual audits中国注册会计师协会2026-03-05Ahttps://cicpa.org.cn/xxfb/news/202603/t20260305_65842.htmlGovernance and Risk
W39《中国注册会计师审计准则问题解答第 18 号 / 第 19 号》中国注册会计师协会2025-01-07Ahttps://cicpa.org.cn/xxfb/news/202501/t20250123_65229.htmlGovernance and Risk
W40Report on the identical adoption of ISO 37301:2021 as GB/T 35770—2022中国标准化研究院2022Ahttps://www.cnis.ac.cn/bydt/zhxw/202210/t20221020_54060.htmlGovernance and Risk
W41Introduction to smart manufacturing capability maturity assessment (GB/T 39116-2020, GB/T 39117-2020)中国电子技术标准化研究院—Ahttps://www.cc.cesi.cn/service/show-2478.aspxIndustry Enablement

3.4. International Governance Frameworks and Industry Standards

No.NameOrganizationYearLevelLinkApplicable sections
W42NIST AI Risk Management Framework (AI RMF 1.0)NIST2023Ahttps://www.nist.gov/itl/ai-risk-management-frameworkGovernance and Risk
W43NIST AI 600-1 Generative Artificial Intelligence ProfileNIST2024Ahttps://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligenceGovernance and Risk
W44ISO/IEC 42001:2023 (Information technology — Artificial intelligence — Management system)ISO/IEC JTC 1/SC 422023Ahttps://www.iso.org/standard/42001Governance and Risk
W45EU AI Act (Regulation (EU) 2024/1689) Implementation TimelineEuropean Commission AI Act Service Desk2024—2026Ahttps://ai-act-service-desk.ec.europa.eu/en/ai-act/timeline/timeline-implementation-eu-ai-actGovernance and Risk
W46SR 11-7 compilation (Navigating Artificial Intelligence in Banking)Bank Policy Institute2024Bhttps://bpi.com/wp-content/uploads/2024/04/Navigating-Artificial-Intelligence-in-Banking.pdfGovernance and Risk
W47HKMA 《Supporting Adoption of Artificial Intelligence in Fighting Financial Crime》香港金融管理局2026-06-22Ahttps://brdr.hkma.gov.hk/eng/doc-ldg/current/20260622-1-ENGovernance and Risk
W48IAASB global roundtable feedback summary on technology and quality managementIAASB2026Ahttps://www.iaasb.org/news-events/2026-02/iaasb-publishes-global-roundtable-feedback-technology-and-quality-managementGovernance and Risk
W49IIA 《Global Internal Audit Standards》The Institute of Internal Auditors2024Ahttps://www.theiia.org/en/standards/documents/Governance and Risk
W50OWASP Top 10 for LLM Applications (2025 edition)OWASP GenAI Security Project2025Ahttps://genai.owasp.org/llm-top-10/Governance and Risk
W51MITRE ATLASMITRE2023—2026Ahttps://atlas.mitre.orgGovernance and Risk

4. Academic Papers

No.NameAuthor / OrganizationYearLevelLinkApplicable sections
W52ReAct: Synergizing Reasoning and Acting in Language ModelsYao et al. (Princeton / Google Brain)2022 (ICLR 2023)Ahttps://arxiv.org/abs/2210.03629Related sections on development history
W53Toolformer: Language Models Can Teach Themselves to Use ToolsSchick et al. (Meta AI)2023 (NeurIPS 2023)Ahttps://arxiv.org/abs/2302.04761Related sections on development history
W54SWE-bench: Can Language Models Resolve Real-World GitHub Issues?Jimenez, Yang et al.2023 (ICLR 2024 Oral)Ahttps://arxiv.org/abs/2310.06770Technical Architecture / Benchmarks
W55Gorilla: Large Language Model Connected with Massive APIsUC Berkeley2023Ahttps://arxiv.org/abs/2305.15334Technical Architecture
W56A Survey of Context Engineering for Large Language ModelsMei et al.2025Bhttps://arxiv.org/abs/2507.13334Technical Architecture / Outlook
W57CyberSentinel-LLM (survey on trust in SOC agents)Tech Science Press (CMC vol.89 no.1)2025Bhttps://www.techscience.com/cmc/v89n1/68397/htmlGovernance and Risk

5. Industry Reports and Surveys

No.NameOrganizationYearLevelLinkApplicable sections
W582025 Stack Overflow Developer SurveyStack Overflow2025-07-29Ahttps://survey.stackoverflow.co/2025/Outlook
W59Stack Overflow 2025 Developer Survey official press releaseStack Overflow2025Ahttps://stackoverflow.co/company/press/archive/stack-overflow-2025-developer-survey/Outlook
W60DORA 2025 State of AI-assisted Software DevelopmentGoogle Cloud / DORA2025-09Ahttps://dora.dev/dora-report-2025 Outlook
W61The State of AI 2025: Agents, Innovation, and TransformationMcKinsey2025-11-05Ahttps://www.mckinsey.com/ Outlook
W62AI Adoption Stats & Trends (aggregation of DORA / McKinsey / JetBrains)daily.dev2026-07-19Bhttps://daily.dev/agentic-ai-hub/ai-adoption-stats-trends/Outlook
W63Cost of a Data Breach Report 2025IBM / Ponemon Institute2025-07-30Ahttps://newsroom.ibm.com/2025-07-30-ibm-report-13-of-organizations-reported-breaches-of-ai-models-or-applications,-97-of-which-reported-lacking-proper-ai-access-controlsGovernance and Risk
W64DARPA AI Cyber Challenge final resultsDARPA2025Ahttps://www.darpa.mil/news/2025/aixcc-resultsGovernance and Risk / Technical Architecture
W65Harness engineering for coding agent usersBirgitta Böckeler / martinfowler.com2026Ahttps://martinfowler.com/articles/harness-engineering.htmlTechnical Architecture / Outlook
W66Humans and Agents in Software Engineering Loopsmartinfowler.com2026Ahttps://martinfowler.com/articles/exploring-gen-ai/humans-and-agents.htmlOutlook
W67My AI Adoption JourneyMitchell Hashimoto2026-02-05Ahttps://mitchellh.com/writing/my-ai-adoption-journeyTechnical Architecture
W68Embracing the wave of intelligence, swimming deep into transformation (新华社 editing assistant, 芒果 large model, 人民网 corpus)新华社2025Ahttps://www.news.cn/20251114/1c01598d836c449bbfd54267b6ecea6d/c.htmlIndustry Enablement
W69Research report on the development of new media run by mainstream media (2024—2025)人民网2025Ahttps://sc.people.com.cn/BIG5/n2/2025/1030/c345167-41396739.htmlIndustry Enablement
W70Will AI give micro-short dramas a “shuffle”? (citing 中国网络视听协会 《微短剧创作指引》)人民日报2026Ahttps://kpzg.people.com.cn/n1/2026/0511/c404214-40717117.htmlIndustry Enablement
W71Premium Times adopts Google NotebookLM to streamline newsroom processesINMA2025Ahttps://www.inma.org/blogs/conference/post.cfm/premium-times-adopts-google-notebook-lm-to-streamline-newsroom-processesIndustry Enablement
W72《技术不是侵权“挡箭牌” 法院这样认定 AI“盗脸”》 (report on the 北京互联网法院 judgment taking effect in 2026-03)新华社《经济参考报》2026-04-17Bhttp://dz.jjckb.cn/www/pages/webpage2009/html/2026-04/17/content_115180.htmGovernance and Risk
W73《e案e审丨短剧角色 AI 换脸“神似”知名演员》北京互联网法院 (contributed) / 澎湃新闻2026Ahttps://www.thepaper.cn/newsDetail_forward_32799628Governance and Risk
W74JPMorgan Chase 2016 annual report (COiN contract intelligence platform)JPMorgan Chase2016Ahttps://reports.jpmorganchase.com/investor-relations/2016/ar-ceo-letter-matt-zames.htmIndustry Enablement
W75A&O Shearman wins FT Innovative Lawyers Awards — official press releaseA&O Shearman (Allen Overy)2024-09-12Ahttps://www.allenovery.com/en/news/ao-shearman-wins-most-innovative-law-firm-in-europe-at-the-ft-innovative-lawyers-awardsIndustry Enablement
W76法信 legal foundation model — official introduction人民法院出版社2024—2025Ahttps://www.faxin.cn/html/about/about.aspxIndustry Enablement
W77中央网信办 report on the scale of generative AI filing users (490+ products / 2.3 亿 users)央视网2025-09Bhttps://big5.cctv.com/gate/big5/news.cctv.cn/2025/09/01/ARTI3ZlXK7MyM39Pm3PuZ5Hm250901.shtmlIndustry Enablement
W782025 comprehensive report on AI applications at listed banks (工银智涌, 邮智, etc.)新华网2025-09-10Bhttps://www.xinhua.org/20250910/4e421fdeeb2242fc97299d73b2d62f09/c.htmlIndustry Enablement
W79Securities firms' 2025 annual-report IT investment statistics (34-firm / 275.9 亿元 account)证券时报2026-04Bhttps://www.stcn.com/article/detail/3731822.htmlIndustry Enablement
W80Securities firms' AI investment research and compliance practices (21 世纪经济报道 feature)21 世纪经济报道2026-04-27Bhttps://www.21jingji.com/article/20260427/herald/b0e4a124fc6808e55600d6fcd453da08.htmlIndustry Enablement
W80aClaude Code Changelog (v2.1.257—269 official version record, 2026-09)Anthropic (official)verified against the 2026-09-13 snapshotAhttps://code.claude.com/docs/en/changelogOutlook
W80bOpenAI API Changelog (2026-09-10 entry: Agents API public beta; GPT-Live 1 GA; 2026-09-03: GPT-6 Astra GA)OpenAI (official)2026-09-10Ahttps://developers.openai.com/api/docs/changelog (2026-09-13 snapshot correction: public beta date corrected from the media-reported 09-11 to the official 09-10)Industry Landscape / Outlook
W80gMCP roadmap (2026-08-22: agent message primitives / HTTP-native transport unification / agent identity)MCP (official)2026-08-22Ahttps://modelcontextprotocol.io/development/roadmapOutlook
W80hCursor Projects (Beta) and Self-Hosted Machines update logCursor (official)2026-09-10 / 09-02Ahttps://cursor.com/changelog (capability descriptions verified against Changelog aggregation-site relays)Outlook
W80iOpenAI research organization agent usage data (3.1 agent workdays per human workday; median researcher spend over USD 600 per day)OpenAI2026-09-06AOfficial figures; original blog post URL [To be verified]Outlook
W80j超维动力 (Kinetix AI) raises over 5 亿元 in an Angel+ round (led by 祥峰资本)中国新闻网 Guangdong channel / 新京报贝壳财经2026-09-11Bhttps://www.gd.chinanews.com.cn/2026/2026-09-11/449585.shtmlOutlook
W80kAccomplish discloses details of a Claude Code macOS sandbox escape (reported 07-13, fixed in v2.1.247)Accomplish2026-09-11BSecurity vendor technical disclosure; original URL [To be verified]Governance and Risk
W80lGemini 3.8 Flash released (HLE-Verified 54.9%, DeepSWE v1.1 approx. 73.8%; $0.75/$3.75 per million tokens; Flash Cyber released simultaneously)Google (official announcement) / Info Data (il Sole 24 Ore)2026-09-02 / 09-12A / BOfficial original URL [To be verified]; details checked against https://www.infodata.ilsole24ore.com/2026/09/12/come-e-fatto-gemini-3-8-flashOutlook
W80mNVIDIA agrees to acquire Hugging Face (approx. 129.3 亿美元; signed 2026-09-02, announced 09-03, expected to close in H1 2027)ECM Source (citing SEC 8-K) / Inside AI2026-09-03A (SEC document relay)https://ecmsource.com/nvidia-hugging-face-12-9-billion-acquisition-september-2026/ (260914 snapshot upgrade: conflict between the two sources closed)Outlook / Industry Landscape
W80nPost-mortem of the OpenAI agent RubyGems abuse incident (occurred 2026-05; RubyGems froze signups for four days)Syntackle (The Weekly Diff #10, third-party investigation summary)early 2026-09Bhttps://syntackle.com/blog/the-ai-slowdown-pact-nvidia-s-13b-hugging-face-deal-the-rubygems-attack-the-weekly-diff-10Governance and Risk
W80oGitHub Copilot 2026-09 updates (HydraFusion, Jira integration, code review ensemble: +47%/+31%/-8%; model retirement on 2026-10-02)GitHub official Changelog (verified against AI Coding Roundup)2026-09-07 / 09-11A / B (effectiveness figures self-reported by the vendor)https://oday-bakkour.com/blog/ai-coding-roundup-september-13-2026Outlook
W80pAnalysis of MCP ToolAnnotations default values (unannotated tools are treated as destructive by default; census of 7 memory servers)deniz.in (third-party analysis); the MCP 2026-07-28 schema original is Level A2026-09-12Bhttps://deniz.in/mcp-s-annotation-defaults-make-unannotated-tools-implicitly-destructiveGovernance and Risk
W80qAWS Kiro student program (free for one year at 132 universities in 18 countries, 1,000 credits per month)AWS (official news)early-to-mid 2026-09 (exact date [To be verified])Ahttps://www.aboutamazon.com/news/aws/amazon-expands-free-kiro-access-students-universitiesOutlook
W80cKiro Web availability and billing basis (confirmed by official FAQ; media-reported GA 2026-09-01)AWS / Kiro (official)verified against the 2026-09-14 snapshotAhttps://kiro.dev/faq (exact GA date [To be verified])Outlook
W80dFigure AI and Nscale compute partnership (initial commitment of 35 亿美元, plans for up to 100,000 Vera Rubin GPUs; Nscale makes a strategic investment in Figure)Embodied AI Applications and Investment Weekly (2026.9.2—9.8)early 2026-09BSupported by an industry weekly; no official announcement found, [To be verified] retained; original URL [To be verified]Outlook
W80f云蝶科技 RoboForge system-level embodied brain (the Harness module records the evidence chain of the entire task process; co-developed with 南洋理工大学; topping the benchmark is vendor-self-reported)南方+ / 云蝶科技 official website2026-09-03—06B + Chttps://www.nfnews.com/content/lyKmjgJ765.html; http://www.cloudbutterfly.com.cn/newsshow/post-227.html (name correction: the original entry misrecorded “云蝶数据”)Outlook

6. Open-Source Projects

No.NameOrganization / AuthorLevelLinkApplicable sections
W81Model Context Protocol specification and SDK (GitHub organization)MCP / AAIFAhttps://github.com/modelcontextprotocolTechnical Architecture
W82OpenAI Codex CLIOpenAIAhttps://github.com/openai/codexTechnical Architecture / Outlook
W83Claude Agent SDKAnthropicAhttps://github.com/anthropics/claude-agent-sdk-pythonTechnical Architecture
W84Google Agent Development Kit (ADK)GoogleAhttps://github.com/google/adk-pythonTechnical Architecture
W8512-Factor AgentsHumanLayer（Dex Horthy）Ahttps://github.com/humanlayer/12-factor-agents Technical Architecture / Outlook
W86Terminal-Bench / Harbor frameworkStanford / Laude Institute / communityAhttps://github.com/harbor-framework/terminal-benchBenchmarks
W87@anthropic-ai/sandbox-runtimeAnthropicAhttps://www.npmjs.com/package/@anthropic-ai/sandbox-runtimeGovernance and Risk
W88ComfyUI (open-source node-based generation workflow)Comfy OrgAhttps://github.com/comfyanonymous/ComfyUIIndustry Enablement

7. Benchmark Evaluations

No.NameOrganizationYearLevelLinkApplicable sections
W89SWE-bench official site and leaderboardPrinceton / community2023—2026Ahttps://www.swebench.com/Technical Architecture / Outlook
W90Terminal-Bench (official site)Stanford / Laude Institute2025—2026Ahttps://www.tbench.ai/ Technical Architecture / Outlook
W91SWE-bench Verified progress timeline 2023—2026 (contains high-impact quantitative figures; must be marked [To be verified])AgentMarketCap2026-04-09Chttps://agentmarketcap.ai/blog/2026/04/09/swe-bench-verified-progress-timeline-2023-2026Technical Architecture
W92Terminal-Bench: The CLI Autonomy Standard (Level C; figures must be marked [To be verified])AgentMarketCap2026-04-09Chttps://agentmarketcap.ai/blog/2026/04/09/terminal-bench-cli-autonomy-standard-coding-agentsTechnical Architecture
W93Artificial Analysis (third-party evaluation leaderboard for image / video generation models)Artificial Analysis2025—2026Bhttps://artificialanalysis.ai/Market Research

8. Official Platform Documentation

No.NameOrganizationYearLevelLinkApplicable sections
W94Claude Code official documentation · SandboxingAnthropic2026Ahttps://code.claude.com/docs/zh-TW/sandboxingGovernance and Risk
W95Claude Code official documentation · Choose a sandbox environmentAnthropic2026Ahttps://code.claude.com/docs/en/sandbox-environmentsGovernance and Risk
W96火山引擎 Ark console — Doubao-Seedance-2.0 series字节跳动 / 火山引擎2026Ahttps://console.volcengine.com/ark/region:ark+cn-beijing/model/detail?Id=doubao-seedance-2-0Market Research
W97即梦 AI official website字节跳动 (剪映 team)2024—2026Ahttps://jimeng.jianying.com/Market Research
W98MCP specification evolution entry (docs.rs mcpkit ProtocolVersion)mcpkit2026Ahttps://docs.rs/mcpkit/latest/enum.ProtocolVersion.htmlTechnical Architecture
W99Model Context Protocol (encyclopedia entry; specification evolution and the three roles)Wikipedia / Klu—Bhttps://en.wikipedia.org/wiki/Model_Context_Protocol; http://klu.ai/glossary/model-context-protocolTechnical Architecture
W100Agent2Agent (A2A) (encyclopedia entry)Wikipedia—Bhttps://en.wikipedia.org/wiki/Agent2AgentOutlook
W101Harness architecture (encyclopedia entry)百度百科2026Bhttps://baike.baidu.com/item/Harness%E6%9E%B6%E6%9E%84/67704948Entire Document
W102Agent Harness: the core paradigm of AI engineering in 2026 (framework-level discussion usable as a lead; figures marked [To be verified])腾讯云开发者社区2026Chttps://developer.cloud.tencent.com/article/2698416Entire Document

9. List of Items Cited in This Whitepaper but Not Authoritatively Confirmed

The following items failed to obtain an A/B-level first-hand source in this project's retrieval, or have conflicts between sources. Where cited in the whitepaper body, they are all annotated in place with [To be verified] or presented side by side; they are aggregated here for subsequent revision and secondary evidence collection.

No.ItemIssueAffected sections
1Publication date of GB/Z 185—2026Three accounts coexist: 2026-05-22 (中国日报) / 2026-06-26 (百度百科) / 2026-07-09 (人民网 “recently released”); the official formal announcement shall prevailOutlook
2GB 45438—2025 《网络安全技术 人工智能生成合成内容标识方法》The standard number and clauses did not obtain an official text link in early retrieval; its metadata fields and explicit labeling specifications come from public relays obtained by this project's market research groupGovernance and Risk
3The specific clause of Article 7 of the 《标识办法》 (verification obligation of application distribution platforms)The original official published text was not directly checkedGovernance and Risk
4GB/T 45654—2025 “the appendix lists 31 categories of risks”, “the compliance pass rate for generated content is no lower than 90%”Secondary interpretation; the original standard text was not verified, and this whitepaper does not adopt itGovernance and Risk
5即梦 AI — the official notice of the 2026-04-28 investigation and subsequent rectificationCompiled from public reports and encyclopedia entries; no administrative punishment decision foundGovernance and Risk
6The complete case number of the AI face-swap judgment that took effect at the 北京互联网法院 in 2026-03Not disclosed in public reportsGovernance and Risk
7The AI Omnibus adjustment to the EU AI Act timeline (2026-05 provisional agreement)A Chinese second-hand source, handled as “proposed”; the EU official service desk shall prevailGovernance and Risk
8The complete list of the 12 risk categories in NIST AI 600-1Another source says 13 categories; the original list was not obtainedGovernance and Risk
9The full title, publication date, and clauses of the 最高人民法院 《关于规范和加强人工智能司法应用的意见》Only a relay was obtainedGovernance and Risk
10The document number and clauses of the 《金融领域科技伦理指引》; the JR/T numbers of the 《人工智能算法金融应用评价规范》 and the 《人工智能算法金融应用信息披露指南》Not obtainedGovernance and Risk
11China's dedicated regulatory texts for medical AI (medical software / assisted-diagnosis review requirements)No clause-level source found; the body text has honestly stated thisGovernance and Risk
12Two statistical accounts for securities firms' IT investment (34 firms / 275.9 亿元 / 6.2% vs. 28 firms / approx. 250 亿元 / 6%)The accounts conflict; this whitepaper adopts the 34-firm account and cites the sourceIndustry Enablement
13All figures and official URLs for DORA 2025, McKinsey 《The State of AI 2025》, and JetBrains 2025All come from reprints by aggregation sites; the official report originals were not verifiedOutlook
14Moves by Chinese vendors (DeepSeek formed a Harness team on 2026-05-20, 小米 released MiMo Code V0.1.0 on 2026-06-11, 灵犀智涌 ROSS in 2026-08)All appear only in encyclopedia-entry relaysOutlook
15The Rust rewrite ratio of Codex CLI (approx. 95% in early 2026)Level B sourceOutlook
16Architecture refactoring case figures (Manus refactored five times in six months, LangChain redesigned three times, Vercel removed 80% of tools)Level B/C sourcesOutlook
17Quantitative figures for context degradation (“performance drops by over 45%”, etc.)Level C source; not confirmed at A/B levelOutlook
18The market size of the Harness layer itselfNo authoritative measurement found; [To be filled]Outlook
19The AGNTCY projectOnly mentioned in surveys; no authoritative first-hand material for nowOutlook
20The ISO/IEC-level international standard for agent interconnectionNo published or approved standard number foundOutlook
21The originator and first appearance of the "Agent Harness" termNo exact first-hand document found; it can be confirmed that Anthropic already used "harness" in 2025 to describe the Claude Agent SDK, and OpenAI pushed "Harness engineering" into the mainstream on 2026-02-11Entire Document
22The 2026 leaderboard figures for SWE-bench Verified and Terminal-Bench 2.0 (W91, W92)Level C source; citations must be marked [To be verified]Technical Architecture
23The Anthropic multi-agent research system blog link (W09) and the 12-Factor Agents repository link (W85)Link accessibility not verified; marked [To be verified]Technical Architecture
24AI risk-control effectiveness metrics in China's banking sector (quantification of bank cases is mostly a relay of securities research, not a verbatim check of annual-report originals)Level BIndustry Enablement
25The original document number of the 中国网络视听协会 《微短剧创作指引》Only a reprint via 人民日报Industry Enablement

Information Gap Statement

In addition to the 25 items in the table above, two overall gaps need to be explained to readers:

  1. Time cross-section: All materials in these references are as of the 2026-09-14 snapshot (the three rounds of incremental retrieval on 2026-09-12/13/14 have been merged into W80a—W80q, information cutoff 2026-09-13). Regulations, standards, and products in the agent field update frequently; before formal citation, the latest version of each source should be rechecked.
  2. Retrieval scope: This whitepaper adopts the “based on existing materials” mode; the materials come from this project's existing research documents and retrieval reports (covering all generated documents of the overview, the eight industry-enablement groups, and the seven market-research groups); no new online retrieval was added; therefore this list only aggregates gaps already identified by existing retrieval, and does not rule out the existence of unidentified gaps.