{
 "generated": "2026-11-05",
 "sha": "707df050bf85",
 "node": {
  "id": "ep:108",
  "type": "Episode",
  "name": "EP.108",
  "slug": "108",
  "ep": 108,
  "date": "2026-10-23",
  "page": "/buzzwords/108/",
  "n_models": 49,
  "window_from": "2026-10-16",
  "window_to": "2026-10-23",
  "window": "2026-10-16 → 2026-10-23",
  "meta": "窗口内发布或开放的主要新模型，按类型分组；每个模型列开发者、发布日、规模、可得性和许可、要点与一手链接。带“边界日”或“背景”标记的，是本期的故事线和快讯里讲到、但发布日在窗口之前的模型。",
  "intro": "发布日以官方公告或模型卡为准，不是 HF 仓库的创建日——仓库往往先建后发，同一个模型的仓库日期可以比公告早一两个月。要点后括号里标了信息来自“厂商自述”“模型卡”还是“独立”：厂商自述的基准本刊没有重跑，模型卡指规格和许可类事实来自 HF 卡本身，独立指第三方榜单或媒体实测。各类按窗口内的在前、边界日和背景的在后排。",
  "excluded": "没有列入的，以及原因。核过、不是窗口内新发布的：Meta Muse Glimmer 30B（8-10 放权重）、Qwen3.8-Flash-Next（8-26 发布，10-22 出现的是社区量化）、阶跃 Step 5 Preview（9-20 放权重，仅作背景；Artificial Analysis 开放权重榜第 4）、Odyssey-3（10-08 研究预览）、Nemotron Parse 2.0（6 月）、Kumo Forecast（4 月）、Speechmatics Linden 1（9-17）、Seedance 2.5（7-31）、Shieldstral 与 Leanstral（7 月）；只是旧模型的修订版：Granite Speech 5.0 的 470M Turbo CTC 与 4.1 2B Plus；不是权重：IBM Granite Timeseries Ensemble R1（一份配方，没有合并权重）。没核到日期或发布稿，暂不列：Bria Fibo 2.0、IBM Granite Switch 4.1（博客写 10-15，模型卡没打开）；也核过但不列：千问的 AI 眼镜 N1 / N1 Pro 和 AI 耳夹式耳机（10-13 发布，在窗口前，且是消费硬件）、火山引擎 Seedance 2.5 4K（10-22 只是向 20 家工作室和导演的定向试用，没有公开价格和可得性，只有媒体报道）、Marathon Fusion 的“汞变金”（公司自述，没有公开数据，也不是模型。",
  "n_in_window": 33,
  "cat_notes": [
   {
    "category": "open-llm",
    "text": "权重能下载的模型。许可证写在每行的“可得性”里：Apache-2.0 和 MIT 是常见的宽许可，Muse Model License、OpenMDW 和 NVIDIA 自家的许可各有条件，别当成同一类。"
   },
   {
    "category": "closed-llm",
    "text": "只能通过 API 或订阅用的模型。Mistral Large 4 官方说是开放权重模型，但权重还没放出，所以也放在这里。"
   },
   {
    "category": "decision",
    "text": "只回答答案有限的问题（选一个、判是非、打分），返回概率，省 token、延迟低。上期说“美国公司的决策模型基座全是 Qwen”，OpenAI 这周入场用的是自家的 GPT-6 Luna，是个反例；中国一侧这周有两个入场者。"
   },
   {
    "category": "science",
    "text": "这一组的基准几乎都是各家自跑的，生物类还多数只有计算验证。"
   },
   {
    "category": "accel",
    "text": "草稿模型要配一个目标模型用，单独跑没有意义。"
   }
  ],
  "title": "思维链连成一坨、新云全靠 Anthropic 吃饭、agent 八成时间在等人点头",
  "oneliner": "出事后开口的人多了，OpenAI 披露页新增 agent 越权访问，但事故上报仍无官方规则；CoreWeave 债券收益率近 16%，AI 的债更贵了；美国有了大块头开放权重，榜单前六仍全是中国模型；专题：Anthropic 的内部就业数据，agent 入口的开与拦。",
  "template": "new",
  "status": "published",
  "n_must_index": 18,
  "n_hl_index": 17,
  "kind": "full",
  "kind_zh": "完整页",
  "entry": "agent",
  "by": "episode-import",
  "at": "2026-11-04T16:15:12",
  "n_stories": 10,
  "n_topics": 7,
  "n_items": 70,
  "n_sections": 6,
  "n_recap": 7,
  "n_must": 13,
  "n_highlight": 3,
  "n_dispatch": 41,
  "n_howto": 1,
  "n_outlook": 5,
  "degree": 0,
  "n_entities": 100,
  "scale": "10 个故事线 / 专题 · 70 个条目 · 讲到 100 个实体"
 },
 "page": "https://xiaopingfeng.com/graph/e/ep/108/",
 "out": [],
 "in": [
  {
   "type": "part_of",
   "sentence": "OpenAI 自己翻出了 agent 越权的事，腾讯也认了，但怎么报、多久报，还没有一条成文的规矩属于EP.108",
   "direction": "in",
   "id": "story:108-story-a",
   "name": "OpenAI 自己翻出了 agent 越权的事，腾讯也认了，但怎么报、多久报，还没有一条成文的规矩",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "CoreWeave 的债券收益率冲到 16%，Anthropic 借钱也要付两位数，数据中心公司的上市接连碰壁属于EP.108",
   "direction": "in",
   "id": "story:108-story-b",
   "name": "CoreWeave 的债券收益率冲到 16%，Anthropic 借钱也要付两位数，数据中心公司的上市接连碰壁",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "美国终于放出了大块头的开源模型，可开放权重榜前六还是全是中国的；同一个基准，配哪个 harness 分数就不一样属于EP.108",
   "direction": "in",
   "id": "story:108-story-c",
   "name": "美国终于放出了大块头的开源模型，可开放权重榜前六还是全是中国的；同一个基准，配哪个 harness 分数就不一样",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Anthropic 拿自己当样本：技术岗里初级的新招占比从 27% 掉到 9%，agent 干的活是员工的 6 倍属于EP.108",
   "direction": "in",
   "id": "story:108-dg-anth-emp",
   "name": "Anthropic 拿自己当样本：技术岗里初级的新招占比从 27% 掉到 9%，agent 干的活是员工的 6 倍",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "agent 被挡在入口外：网站、银行、国内的 App，各有各的拦法，没有一条共同的规则；这周也有人开了门属于EP.108",
   "direction": "in",
   "id": "story:108-dg-agent-gate",
   "name": "agent 被挡在入口外：网站、银行、国内的 App，各有各的拦法，没有一条共同的规则；这周也有人开了门",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "AI 进实验室这一周：设计越来越快，卡点挪到了湿实验和验证上属于EP.108",
   "direction": "in",
   "id": "story:108-dg-ai-bio",
   "name": "AI 进实验室这一周：设计越来越快，卡点挪到了湿实验和验证上",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "DeepSeek 把推理引擎开源了：预填充、解码、KV 缓存拆成三块，各自是怎么省钱的属于EP.108",
   "direction": "in",
   "id": "story:108-dg-deepinfer",
   "name": "DeepSeek 把推理引擎开源了：预填充、解码、KV 缓存拆成三块，各自是怎么省钱的",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "用 agent 做数据管线：两种用法，一个共同的办法，每一步都要有能打分的检查属于EP.108",
   "direction": "in",
   "id": "story:108-dg-agent-pipeline",
   "name": "用 agent 做数据管线：两种用法，一个共同的办法，每一步都要有能打分的检查",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Agent 开始卖货给 agent：卖家只被告知“多被选中”，五个模型都自发把描述越写越满；一个模型下，买家退回随机属于EP.108",
   "direction": "in",
   "id": "story:108-dg-agent-sellers",
   "name": "Agent 开始卖货给 agent：卖家只被告知“多被选中”，五个模型都自发把描述越写越满；一个模型下，买家退回随机",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "模型可解释性这三个月：Goodfire 说神经元本来就干净，同行说证据还不够；真正在上线的是拿内部信号做的探针属于EP.108",
   "direction": "in",
   "id": "story:108-dg-interp",
   "name": "模型可解释性这三个月：Goodfire 说神经元本来就干净，同行说证据还不够；真正在上线的是拿内部信号做的探针",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Anthropic 的公开 S-1，会在 10 月 23 号前出现在 EDGAR 上属于EP.108",
   "direction": "in",
   "id": "item:108-recap-1",
   "name": "Anthropic 的公开 S-1，会在 10 月 23 号前出现在 EDGAR 上",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "白宫的“不可选”：10 月 23 号前，白宫不会发布写明事故门槛、上报期限或处罚的书面文件属于EP.108",
   "direction": "in",
   "id": "item:108-recap-2",
   "name": "白宫的“不可选”：10 月 23 号前，白宫不会发布写明事故门槛、上报期限或处罚的书面文件",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "OpenAI 说数周内公布的第三方评估合同：10 月 23 号前不会公布任何一家具名的评估方合同属于EP.108",
   "direction": "in",
   "id": "item:108-recap-3",
   "name": "OpenAI 说数周内公布的第三方评估合同：10 月 23 号前不会公布任何一家具名的评估方合同",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Reflection Beam 和 Mistral Large 4 的权重：10 月 23 号前，Reflection 不会放出 Beam 的权重，Mistral 也不会放出 Large 4 的权重属于EP.108",
   "direction": "in",
   "id": "item:108-recap-4",
   "name": "Reflection Beam 和 Mistral Large 4 的权重：10 月 23 号前，Reflection 不会放出 Beam 的权重，Mistral 也不会放出 Large 4 的权重",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Lean 官方渠道：10 月 23 号前不会公布新的、被确认能用来证明 False 的内核或运行时漏洞属于EP.108",
   "direction": "in",
   "id": "item:108-recap-5",
   "name": "Lean 官方渠道：10 月 23 号前不会公布新的、被确认能用来证明 False 的内核或运行时漏洞",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "FTC 对 Anthropic 和 OpenAI 的调查令：10 月 23 号前，FTC 不会就这批调查令发布官网新闻稿，也不会提起诉讼或发布执法令属于EP.108",
   "direction": "in",
   "id": "item:108-recap-6",
   "name": "FTC 对 Anthropic 和 OpenAI 的调查令：10 月 23 号前，FTC 不会就这批调查令发布官网新闻稿，也不会提起诉讼或发布执法令",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "批量生成的报警和线索：10 月 23 号前，FBI、国土安全部、CISA 或任一州的警察和应急机构，不会发布专门点名“生成式 AI 批量生成的报警、线索或 911 来电”的通告或指引属于EP.108",
   "direction": "in",
   "id": "item:108-recap-7",
   "name": "批量生成的报警和线索：10 月 23 号前，FBI、国土安全部、CISA 或任一州的警察和应急机构，不会发布专门点名“生成式 AI 批量生成的报警、线索或 911 来电”的通告或指引",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Claude reveals ECDSA curve origins属于EP.108",
   "direction": "in",
   "id": "item:108-hl-ecdsa",
   "name": "Claude reveals ECDSA curve origins",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Designing more efficient quantum circuits with Claude属于EP.108",
   "direction": "in",
   "id": "item:108-hl-quantum",
   "name": "Designing more efficient quantum circuits with Claude",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Postgres 19 delayed after AI reviewers find more bugs than humans can fix属于EP.108",
   "direction": "in",
   "id": "item:108-hl-pg19",
   "name": "Postgres 19 delayed after AI reviewers find more bugs than humans can fix",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Advancing coding in low-resource programming languages属于EP.108",
   "direction": "in",
   "id": "item:108-hl-lowres-lang",
   "name": "Advancing coding in low-resource programming languages",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Reducing Reward Hacking by Localizing Bad Behavior Instead of Penalizing It属于EP.108",
   "direction": "in",
   "id": "item:108-hl-graft",
   "name": "Reducing Reward Hacking by Localizing Bad Behavior Instead of Penalizing It",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "LongZero: Enhancing Long-Context Reasoning in LLMs from Zero Data属于EP.108",
   "direction": "in",
   "id": "item:108-hl-longzero",
   "name": "LongZero: Enhancing Long-Context Reasoning in LLMs from Zero Data",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "LLMs are making formal software verification feasible at a much larger scale属于EP.108",
   "direction": "in",
   "id": "item:108-hl-formal",
   "name": "LLMs are making formal software verification feasible at a much larger scale",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "A team of Gemini 4 Argon agents beats Minecraft属于EP.108",
   "direction": "in",
   "id": "item:108-hl-minecraft",
   "name": "A team of Gemini 4 Argon agents beats Minecraft",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "PHP 8.5.12 / 8.4.27 / 8.3.36 / 8.2.35 安全发布属于EP.108",
   "direction": "in",
   "id": "item:108-hl-php",
   "name": "PHP 8.5.12 / 8.4.27 / 8.3.36 / 8.2.35 安全发布",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "How one hacker used AI agents and a few thousand dollars to break into scores of sites属于EP.108",
   "direction": "in",
   "id": "item:108-hl-norforge",
   "name": "How one hacker used AI agents and a few thousand dollars to break into scores of sites",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Inside CARLIY0S: How CyberCartel Used Multiple AI Models to Attack 400+ Latin American Banks and Financial Institutions属于EP.108",
   "direction": "in",
   "id": "item:108-hl-cybercartel",
   "name": "Inside CARLIY0S: How CyberCartel Used Multiple AI Models to Attack 400+ Latin American Banks and Financial Institutions",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "LiteLLM 1.107.0 在 PyPI 被投毒属于EP.108",
   "direction": "in",
   "id": "item:108-hl-litellm",
   "name": "LiteLLM 1.107.0 在 PyPI 被投毒",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "eMudhra 的 8 张中间 CA 证书，私钥可以从证书里直接还原属于EP.108",
   "direction": "in",
   "id": "item:108-hl-emudhra",
   "name": "eMudhra 的 8 张中间 CA 证书，私钥可以从证书里直接还原",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Eiliko maker says it’s hiring an ‘independent security firm’ after Verge report on its insecure robots属于EP.108",
   "direction": "in",
   "id": "item:108-hl-eiliko",
   "name": "Eiliko maker says it’s hiring an ‘independent security firm’ after Verge report on its insecure robots",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Timestral：Mistral 的时间序列基础模型属于EP.108",
   "direction": "in",
   "id": "item:108-hl-timestral",
   "name": "Timestral：Mistral 的时间序列基础模型",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Introducing Claude Duet属于EP.108",
   "direction": "in",
   "id": "item:108-hl-duet",
   "name": "Introducing Claude Duet",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "humans& 发布 Ampersand属于EP.108",
   "direction": "in",
   "id": "item:108-hl-humansand",
   "name": "humans& 发布 Ampersand",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Voxtral 3属于EP.108",
   "direction": "in",
   "id": "item:108-hl-voxtral3",
   "name": "Voxtral 3",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Claude API 新版本 2026-10-21属于EP.108",
   "direction": "in",
   "id": "item:108-hl-api-1021",
   "name": "Claude API 新版本 2026-10-21",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Claude Platform is now available on Google Cloud属于EP.108",
   "direction": "in",
   "id": "item:108-hl-claude-gcp",
   "name": "Claude Platform is now available on Google Cloud",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Donating Agent Skills to the Agentic AI Foundation属于EP.108",
   "direction": "in",
   "id": "item:108-hl-skills-aaif",
   "name": "Donating Agent Skills to the Agentic AI Foundation",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "SemiAnalysis：Anthropic 在 A/B 测试用量额度属于EP.108",
   "direction": "in",
   "id": "item:108-hl-semi-limits",
   "name": "SemiAnalysis：Anthropic 在 A/B 测试用量额度",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "NVIDIA DSX POD属于EP.108",
   "direction": "in",
   "id": "item:108-hl-dsxpod",
   "name": "NVIDIA DSX POD",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "全球算力价格指数启动属于EP.108",
   "direction": "in",
   "id": "item:108-hl-compute-index",
   "name": "全球算力价格指数启动",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "SpaceX 的第三季度属于EP.108",
   "direction": "in",
   "id": "item:108-hl-spacex-q3",
   "name": "SpaceX 的第三季度",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Flow Engineering 完成 B 轮：5000 万美元，估值 7.5 亿美元属于EP.108",
   "direction": "in",
   "id": "item:108-hl-flow-eng",
   "name": "Flow Engineering 完成 B 轮：5000 万美元，估值 7.5 亿美元",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Investing in the public benefit属于EP.108",
   "direction": "in",
   "id": "item:108-hl-pbf",
   "name": "Investing in the public benefit",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Disney agrees to hire the majority of Character.AI’s workforce属于EP.108",
   "direction": "in",
   "id": "item:108-hl-charai",
   "name": "Disney agrees to hire the majority of Character.AI’s workforce",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "OpenAI agrees not to use alleged Apple trade secrets amid ongoing lawsuit属于EP.108",
   "direction": "in",
   "id": "item:108-hl-apple-liu",
   "name": "OpenAI agrees not to use alleged Apple trade secrets amid ongoing lawsuit",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Introducing Securing America’s Infrastructure Foundation属于EP.108",
   "direction": "in",
   "id": "item:108-hl-saif",
   "name": "Introducing Securing America’s Infrastructure Foundation",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Barclays to Cut 5,000 Jobs as AI Drives Cost Reduction Effort属于EP.108",
   "direction": "in",
   "id": "item:108-hl-barclays",
   "name": "Barclays to Cut 5,000 Jobs as AI Drives Cost Reduction Effort",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "What FDE means at OpenAI属于EP.108",
   "direction": "in",
   "id": "item:108-hl-fde",
   "name": "What FDE means at OpenAI",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Meta’s new terms may allow it to train AI on news sites for free属于EP.108",
   "direction": "in",
   "id": "item:108-hl-meta-terms",
   "name": "Meta’s new terms may allow it to train AI on news sites for free",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Appeals court upholds Thomson Reuters win in AI training lawsuit属于EP.108",
   "direction": "in",
   "id": "item:108-hl-tr-ross",
   "name": "Appeals court upholds Thomson Reuters win in AI training lawsuit",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "OpenAI's 'Pacing 2.0' Problem: How Smart Is Too Smart?属于EP.108",
   "direction": "in",
   "id": "item:108-hl-pacing2",
   "name": "OpenAI's 'Pacing 2.0' Problem: How Smart Is Too Smart?",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Anthropic 2026 Public Benefit Report属于EP.108",
   "direction": "in",
   "id": "item:108-hl-pbr",
   "name": "Anthropic 2026 Public Benefit Report",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Jensen Huang 对“放慢”的回答属于EP.108",
   "direction": "in",
   "id": "item:108-hl-jensen",
   "name": "Jensen Huang 对“放慢”的回答",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "The AI expectations gap属于EP.108",
   "direction": "in",
   "id": "item:108-hl-gap",
   "name": "The AI expectations gap",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "The Top 100 Gen AI Consumer Apps, 7th Edition属于EP.108",
   "direction": "in",
   "id": "item:108-hl-a16z7",
   "name": "The Top 100 Gen AI Consumer Apps, 7th Edition",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "How Chinese companies talk about recursive self-improvement属于EP.108",
   "direction": "in",
   "id": "item:108-hl-asterisk",
   "name": "How Chinese companies talk about recursive self-improvement",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Hassabis 的 WSJ Magazine 人物专访属于EP.108",
   "direction": "in",
   "id": "item:108-hl-hassabis",
   "name": "Hassabis 的 WSJ Magazine 人物专访",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "DeepSeek 搜索与抓取 API：给你的 agent 接一个搜索和网页抓取工具↗属于EP.108",
   "direction": "in",
   "id": "item:108-ht-c4972b",
   "name": "DeepSeek 搜索与抓取 API：给你的 agent 接一个搜索和网页抓取工具↗",
   "status": "derived",
   "evidence": []
  }
 ]
}