AiX 環球 AI 情報 Global AI Intelligence

第 25 期 2026 年 10 月 22 日（星期四） 07:30 出刊 10 條情報 · 15 原文來源連結 RSS 過往期數

今日十條 · 2026 年 10 月 22 日代理開始自己開接口、自己收錢、自己按快門：商業化跑到治理前面的一天

今日主編判斷

今日主線是「代理自己開門」：沃爾瑪以 MCP 網關向代理開放交易並設代理檔案與消費上限；Anthropic 發布主動式代理應用 Duet，同時換上新版 API 與 Apache 2.0 授權，把工程接線變成企業成本；百度把智能體推上生產線。另一邊，Anthropic 推銷的安全法案門檻被指過高。Today's through-line is that agents are opening their own doors. Walmart is opening transactions to agents through an MCP gateway with agent profiles and spending caps; Anthropic has released Duet, a proactive agent app, alongside a new API version and an Apache 2.0 licence that turn integration work into enterprise cost; and Baidu is pushing agents onto the production line. Meanwhile, the risk threshold in the safety bill Anthropic is shopping is criticised as too high.

今日十條

本欄由 AIX Society 編輯部整理公開資料後撰稿，每條均附原始來源連結；事實與觀點分列，觀點屬本會判斷。

按受眾篩選

沃爾瑪向人工智能代理開放購物，設代理檔案與消費上限Walmart opens shopping to AI agents with agent profiles and spending limits

影響力✓ 已核實Verified已在發生Happening now企業決策者Business owners前線員工Frontline staff資訊科技IT法務合規Legal & compliance

事件: 沃爾瑪十月二十一日宣布推出智能體網關，以 MCP 協議讓經審核代理查詢庫存並代客下單，代理檔案下月推出。Walmart said on 21 October that it is launching an Agent Gateway, an MCP server that lets approved AI agents search inventory and buy on a customer's behalf, together with terms for agents; Agent Profiles follow next month.
背後: 亞馬遜近期封鎖外部代理，旅遊與預訂網站亦拒絕代理訪問；沃爾瑪選擇接入，以規則換取代理帶來的訂單。Amazon has recently blocked outside agents, and travel and booking sites have been shutting them out too. Walmart has instead integrated agents such as Muse, ChatGPT and Gemini, trading clear rules for the traffic and orders those agents bring.

顧問觀點

我們認為，重點不在沃爾瑪，而在它的代理條款：不得超出用戶意圖、不得接觸支付資料、違規即封鎖。零售業首次把代理當成有責任的交易對手。We believe the value of this news lies less in Walmart than in the terms it has written for agents: no action beyond shopper intent, no access to payment data, blocked on breach. This is the first time a retailer has managed agents as counterparties with defined responsibilities.

對日常工作的影響: 前線客服要分辨人手下的單與代理下的單，退貨與爭議按不同記錄處理；採購要評估代理訂單的質量。Frontline service staff will have to tell human orders from agent orders and handle returns and disputes against separate records, while procurement and marketing assess the quality and return rate of agent-driven orders.
如何改善: 本週把網站與下單流程的代理政策寫成一頁：准不准、用哪個接口、違規如何封鎖，交法務與資訊部門聯簽。This week, write a one-page agent policy for your website and ordering flow: whether agents are admitted, through which interface, and how breaches are blocked, then have legal and IT co-sign it.

來源The Verge（2026-10-21）↗Unite.AI（2026-10-21）↗

Anthropic 發布主動式代理應用 Claude Duet，先以研究預覽開放美加 Pro 與 Max 用戶Anthropic releases Claude Duet, a proactive agent app, first as a research preview for Pro and Max users in the US and Canada

影響力✓ 已核實Verified已在發生Happening now企業決策者Business owners資訊科技IT人力資源HR

事件: Anthropic 十月二十一日開放 Claude Duet 研究預覽，以 Opus 5.5 為底座，代理執行任務前須用戶批准。Anthropic opened a research preview of Claude Duet on 21 October. Built on Opus 5.5, it works across connected apps in the background and asks for approval before acting, initially for Pro and Max subscribers in the US and Canada who join a waitlist.
背後: 過去代理多由用戶逐步下指令，Duet 改為主動推送簡報；Anthropic 稱憑證由系統代持。Agents over the past year have mostly waited for step-by-step instructions. Duet instead pushes a morning brief and messages when something needs attention. Anthropic says it reuses its existing agent framework, with credentials held by the system so the agent never sees passwords.

顧問觀點

我們認為，主動式代理改變的是誰決定優先次序。當代理自行判斷什麼值得通知，注意力分配就成為須設定與審核的參數。We believe what a proactive agent changes is who sets priorities. Once the agent decides what deserves your attention, attention allocation becomes a parameter that must be configured and reviewed, not a personal habit.

對日常工作的影響: 老闆與中層要界定哪些事項可由代理代為排期、哪些須人先看；資訊部門須發獨立憑證與審計軌跡。Owners and middle managers must redefine which items an agent may schedule on their behalf and which a person must see first, while IT has to issue separate credentials and audit trails rather than sharing personal accounts.
如何改善: 本週請一位主管試用一週，記錄代理推送的通知有多少真正需人處理，據此訂出兩張清單。This week, have one manager trial it for a week and log how many agent notifications genuinely needed a person, then turn that into two lists: what may be scheduled automatically and what needs human review first.

來源Anthropic Release Notes（經 Releasebot 轉載）（2026-10-21）↗

Anthropic 推出 API 版本 2026-10-21，SDK 授權由 MIT 改為 Apache 2.0Anthropic ships API version 2026-10-21, moving SDK licensing from MIT to Apache 2.0

影響力✓ 已核實Verified已在發生Happening now資訊科技IT法務合規Legal & compliance企業決策者Business owners

事件: Anthropic 十月二十一日發布新版 API 2026-10-21 及七種語言 SDK，改用 Apache 2.0 授權。On 21 October Anthropic released API version 2026-10-21 with major-version SDKs for seven languages and a CLI that default to it. The SDK licence moved from MIT to Apache 2.0.
背後: 新版移除 temperature、top_p 與 top_k，傳送即回 400 錯誤；舊版可用但須逐項對照。The new version removes the temperature, top_p and top_k sampling parameters, with requests carrying them returning a 400 error. Several management API fields are renamed. Version 2023-06-01 remains available, but upgrades must be checked item by item.

顧問觀點

我們認為，最易低估的是這類「不是模型的變更」。模型換代可靠評測決定，接口與授權變更卻會令線上服務靜靜失效。We believe enterprises most often underestimate changes that are not about the model. A model upgrade can be decided by evaluation, but interface and licence changes can quietly break live services, and licence terms go straight to legal and procurement.

對日常工作的影響: 資訊部門要鎖定生產環境 SDK 版本並排升級演練；法務要重審授權合約中關於再分發的條款。IT must pin SDK versions in production and schedule an upgrade rehearsal, while legal reviews redistribution clauses in licence agreements, especially for self-hosted products.
如何改善: 本週在測試環境以新版 SDK 跑一次回歸測試，列出哪些請求因取樣參數被拒，再決定去留。This week, run one regression test on the new SDK in staging, list which requests were rejected over sampling parameters, then decide whether to stay on the old version or rewrite the prompting strategy.

來源Wain（工程分析）（2026-10-21）↗Anthropic Release Notes（經 Releasebot 轉載）（2026-10-21）↗

OpenAI 收購藥物研發協調平台 Orchestra Bio，技術併入生命科學產品線 RosalindOpenAI acquires drug development coordination platform Orchestra Bio and folds it into Rosalind

影響力✓ 已核實Verified6–12 個月6–12 months企業決策者Business owners資訊科技IT法務合規Legal & compliance

事件: OpenAI 十月二十日宣布收購 2022 年成立的 Orchestra Bio，併入 Rosalind 產品線。OpenAI said on 20 October that it had acquired Orchestra Bio, founded in 2022, folding its team and technology into Rosalind, its life sciences line. The price was not disclosed.
背後: OpenAI 稱在美研製新藥由靶點到獲批平均需十至十五年；Rosalind 缺的正是排期協調。OpenAI says bringing a new medicine from target to US approval averages ten to fifteen years, and a failed batch can delay a trial. Rosalind covers hypotheses and experiments; what it lacked was scheduling coordination.

顧問觀點

我們認為，這宗收購說明通用模型落地垂直行業的瓶頸不在智能，而在流程。誰能讓代理讀懂節點依賴與合規文件，誰才拿得到高容錯行業的入場券。We believe this acquisition shows the bottleneck for general models entering regulated industries is not intelligence but process. Whoever lets agents read node dependencies and compliance documents holds the entry ticket to low-tolerance industries.

對日常工作的影響: 製藥與醫療器械公司的項目經理最先受影響：延期預警與跨方對賬由代理匯總，人手集中處理判斷與對外協商。Project managers in pharma and medical devices feel this first: delay alerts and cross-party reconciliation are summarised by agents, leaving people for judgement and external negotiation.
如何改善: 本週挑一個已延誤的項目，讓代理重演一次排期影響分析，對比人手結果，記錄兩者差異再決定是否擴大使用。This week, take one already-delayed project and have an agent redo the schedule impact analysis, compare it with the human result, and record the differences before widening its use.

來源The Hindu BusinessLine（2026-10-21）↗IT之家（騰訊新聞轉載）（2026-10-21）↗

Anthropic 向國會推銷的 AI 安全草案設極高門檻：五十人死亡或十億美元損失Anthropic's draft AI safety bill sets a high bar: fifty deaths or one billion dollars in damage

影響力✓ 已核實Verified6–12 個月6–12 months法務合規Legal & compliance企業決策者Business owners

事件: Politico 十月二十一日報道，Anthropic 草案規定模型須構成五十人以上死亡或十億美元損失，法院才可禁發布。Politico reported on 21 October that Anthropic's draft bill would let a court block a model release only where it risks more than fifty deaths or one billion dollars in damage.
背後: 該門檻沿用加州 SB 53 的災難性風險定義，用途卻由披露與評估改為阻止發布，並排除部分州級上報法例。The threshold borrows California SB 53's definition of catastrophic risk but repurposes it from disclosure and assessment to blocking a release. The draft also pre-empts some state rules on risk disclosure and incident reporting, which critics call contradictory.

顧問觀點

我們認為，這條消息對企業的意義是「合規下限將由模型商游說決定」。當門檻定得這樣高，企業不能假設監管會替你攔住高風險場景，只能自建內部紅線。We believe the point for enterprises is that the compliance floor will be set by vendor lobbying. With a threshold this high, companies cannot assume regulators will stop high-risk uses for them and must set their own red lines.

對日常工作的影響: 法務與風險部門要檢視「等監管再說」的假設，醫療、金融與關鍵基建客戶須自行界定不可接受的失敗情景。Legal and risk teams must revisit the assumption of waiting for regulation, especially those serving healthcare, finance and critical infrastructure, and define unacceptable failure scenarios themselves.
如何改善: 本週由法務牽頭寫一頁「本公司不可接受的代理失敗情景」，列出三至五項，並指明每項的停止權由誰行使。This week, have legal draft a one-page list of agent failure scenarios the company will not accept, three to five items, naming who holds the stop authority for each.

來源POLITICO（2026-10-21）↗NPR（2026-10-21）↗

英國公布數碼與科技職位計劃，十年開約二十萬個入門職位UK publishes Digital and Technologies Jobs Plan targeting about 200,000 entry-level roles over a decade

影響力✓ 已核實Verified3 年3 years人力資源HR企業決策者Business owners

事件: 英國政府十月公布《數碼與科技職位計劃》，指未來十年需填補約八十萬個職位，擬提供約二十萬個入門職位。The UK government published its Digital and Technologies Jobs Plan in October, projecting around 800,000 digital and technology roles to fill over the next decade and aiming to open about 200,000 entry-level jobs.
背後: 計劃把見習、T Level 實習與入門招聘整合為單一技能協議，並提前一年於二〇二八年前接觸百萬名年輕人。The plan bundles apprenticeships, T Level placements and entry recruitment into one skills deal, and brings TechFirst forward a year to reach one million young people by 2028.

顧問觀點

我們認為，以補貼職位數回應 AI 對入門工作的擠壓，方向對但見效慢；企業缺的是把新人訓練成會驗收代理輸出的工序。We believe subsidising the number of entry-level roles is the right direction but slow to bite. What firms actually lack is process design that trains newcomers to check agent output, not more places.

對日常工作的影響: 人力資源要重寫見習與初級職位考核標準，加入代理輸出核驗與異常上報，否則新名額只成廉價勞動。HR must rewrite assessment criteria for apprentices and juniors to include checking agent output and escalating anomalies, otherwise the new places become cheap labour.
如何改善: 本週選一個入門崗位，把它第一季的培訓內容加入一項「代理輸出抽查」實作練習，並指定導師評分。This week, pick one entry-level role, add a hands-on agent output sampling exercise to its first-quarter training, and name a mentor to mark it.

來源GOV.UK（2026-10-21）↗

全球最大主權基金淡馬錫等報告：東南亞上半年八成私人融資流向人工智能，企業採用仍落後Temasek, Google and Bain report: four-fifths of Southeast Asia's private funding went to AI in the first half, yet enterprise adoption lags

影響力✓ 已核實Verified6–12 個月6–12 months企業決策者Business owners市場推廣Marketing

事件: 谷歌、淡馬錫與貝恩十月二十一日發表報告，指東南亞上半年流向人工智能應用的資金達三十二億美元。Google, Temasek and Bain said on 21 October that 3.2 billion US dollars of Southeast Asian private funding went to AI applications in the first half, about four-fifths of the total.
背後: 報告指上半年區內數據中心私人融資達一百六十三億美元；但六國人均 AI 用量僅約美國的百分之八。The report puts regional data-centre private funding at 16.3 billion US dollars in the first half, yet per-capita AI use across six countries is about eight per cent of the US.

顧問觀點

我們認為，資金與採用之間的落差，正是顧問服務的市場空間。錢已流入基建與應用層，但企業沒有把代理接進流程，本地化落地與治理配套成為瓶頸。We believe the gap between money and adoption is precisely the consulting market. Capital has reached infrastructure and applications, but firms have not wired agents into processes, making local deployment and governance the bottleneck.

對日常工作的影響: 區域總部與市場部門要回答錢進了哪裡、流程有否跟上；財務要審視 AI 投入是否只買席位而未改工序。Regional headquarters and marketing must answer where the money went and whether their processes kept pace, while finance checks whether AI spend bought seats without changing any process.
如何改善: 本週用一小時抽查公司三個核心流程，記錄哪些步驟已由代理執行、哪些仍全人手，據此列出下季三個試點。This week, spend an hour auditing three core processes, noting which steps an agent already runs and which are fully manual, then list three pilots for next quarter.

來源聯合早報（2026-10-21）↗Fintech News Singapore（2026-10-21）↗

日本召開跨部門聯絡會議加快引進 AI，年內彙總行動計劃Japan convenes cross-ministry meeting to speed AI adoption and will finalise an action plan this year

影響力✓ 已核實Verified6–12 個月6–12 months企業決策者Business owners法務合規Legal & compliance資訊科技IT

事件: 日本政府十月二十一日首次召開府省廳聯絡會議，討論在缺工領域推進 AI，年內彙總行動計劃。Japan's government held the first meeting of a cross-ministry council on 21 October to discuss AI adoption in labour-short sectors such as healthcare, nursing and transport, with an action plan due this year.
背後: 會議由 AX 社會實裝擔當相赤澤亮正主持，方針為以 AI 帶動社會變革，並推進物理 AI。Trade minister Ryosei Akazawa, who holds the AX portfolio, chaired the meeting; the policy uses manufacturing and service data to advance physical AI.

顧問觀點

我們認為，日本這一步的關鍵是把 AI 當作人口問題的解法而非產業政策，府省廳聯席意味著跨部門審批與數據共享要一併鬆綁，難點在執行而非宣言。We believe the key here is that Japan treats AI as an answer to demography rather than only industrial policy. A cross-ministry council means approvals and data sharing must loosen together; execution, not declaration, is the test.

對日常工作的影響: 在日營運企業要留意醫療、護理與運輸的許可與補助變化，法務與營運要追蹤各府省廳實施細則。Firms operating in Japan should watch for changes to licensing and subsidies in healthcare, nursing and transport, with legal and operations tracking each ministry's implementing rules.
如何改善: 本週指定一人訂閱經產省與厚生勞動省的相關公告，建立一頁「日本 AI 政策變動」追蹤表，每月更新一次。This week, assign one person to follow METI and health ministry announcements and set up a one-page Japan AI policy tracker, updated monthly.

來源共同社（2026-10-21）↗

南韓與日本舉行第二次人工智能工作層會議，商討安全治理與知識產權合作South Korea and Japan hold second AI working-level meeting on safety governance and intellectual property

影響力✓ 已核實Verified6–12 個月6–12 months法務合規Legal & compliance企業決策者Business owners

事件: 南韓外交部十月二十一日主持第二次韓日人工智能工作層會議，兩國逾三十名官員出席，討論安全治理與標準。South Korea's foreign ministry hosted the second Korea-Japan AI working-level meeting on 21 October, with more than thirty officials discussing safety governance and standards.
背後: 會議源自去年九月韓日峰會的後續措施，首次會議於今年二月在東京舉行。The meeting follows up on last September's summit; the first session was in Tokyo in February, with Korea's science ministry and other agencies taking part.

顧問觀點

我們認為，日韓把 AI 治理放在雙邊議程而非多邊談判，說明區域標準正在先行成形；企業若同時在兩地營運，可預期評估與申報要求逐步靠攏。We believe Japan and Korea putting AI governance on a bilateral agenda rather than in multilateral talks shows regional standards forming first. Firms operating in both markets can expect evaluation and reporting requirements to converge.

對日常工作的影響: 合規與產品部門要開始比對日韓兩地的模型評估與事故上報要求，找出差異點，避免同一產品要交兩套文件。Compliance and product teams should start comparing model evaluation and incident reporting rules in Japan and Korea, identifying divergences so one product does not need two sets of filings.
如何改善: 本週把日韓兩地的申報要求整理成一張對照表，標出可共用的文件與必須分開準備的部分。This week, turn Japanese and Korean filing requirements into one comparison table, marking documents that can be shared and parts that must be prepared separately.

來源Aju Press（2026-10-21）↗

百度搭子發布企業級智能體能力，內地智能體之爭轉向生產流程Baidu's Dazi ships enterprise agent capabilities as China's agent race shifts to production workflows

影響力✓ 已核實Verified已在發生Happening now資訊科技IT企業決策者Business owners前線員工Frontline staff

事件: 百度十月二十一日在杭州發布百度搭子新能力，涵蓋操作無接口舊系統與 Agent 安全中心。Baidu unveiled new Dazi capabilities in Hangzhou on 21 October, including operating legacy systems without APIs, more than thirty expert kits and an agent security centre.
背後: 內地頭部廠商今年把辦公智能體升為戰略入口，四家大廠均已完成布局，競爭轉向誰能讓代理穩定運行。China's largest platforms made office agents a strategic entry point this year, with four major vendors now positioned and competition shifting to reliable agent operation.

顧問觀點

我們認為，內地競爭已由模型能力轉為接入舊系統的能力。企業核心流程仍留在沒有接口的老系統，誰能安全操作，誰才進得了生產線。We believe competition in China has shifted from model capability to the ability to reach legacy systems. Most core processes still sit on systems without APIs, so whoever can operate them safely is the one that truly reaches the production line.

對日常工作的影響: 資訊部門要為代理操作舊系統設沙箱與授權層級；業務部門要挑出人手重複操作舊系統最嚴重的工序試點。IT must set sandboxing and authorisation tiers for agents operating legacy systems, while business units pick the three workflows where people most repeat manual work in old systems as first pilots.
如何改善: 本週列出仍靠人手操作的舊系統清單，標明每項的數據敏感度與錯誤代價，據此定代理接入次序。This week, list the legacy systems your staff still operate by hand, mark data sensitivity and cost of error for each, and use that to set the priority order for agent access.

來源證券時報（騰訊新聞轉載）（2026-10-22）↗國際金融報（2026-10-21）↗

本週可做的三件事

為每個對外代理設一頁「授權卡」：它能動用多少錢、可改哪些欄位、誰批准超額，寫明預設逐筆批准。Give every outward-facing agent a one-page authorisation card: how much it may spend, which fields it may change, who approves overruns, with per-transaction approval as the default.
盤點生產環境的模型接口版本與 SDK，鎖定版本號並排一次升級演練，記錄哪些參數已被移除。Inventory the model API versions and SDKs running in production, pin the version numbers, run one upgrade rehearsal, and record which parameters have been removed.
挑一條對外流程，寫明「代理出錯時幾小時內可回復、由誰回復」，未寫完前不開放代理自主下單。Pick one outward-facing workflow and write down how quickly an agent error can be reversed and by whom; do not let agents order autonomously until that is written.

常見問題

今日（2026-10-22）「AiX 環球 AI 情報」有哪十條重點？

今日十條為：1、沃爾瑪向人工智能代理開放購物，設代理檔案與消費上限；2、Anthropic 發布主動式代理應用 Claude Duet，先以研究預覽開放美加 Pro 與 Max 用戶；3、Anthropic 推出 API 版本 2026-10-21，SDK 授權由 MIT 改為 Apache 2.0；4、OpenAI 收購藥物研發協調平台 Orchestra Bio，技術併入生命科學產品線 Rosalind；5、Anthropic 向國會推銷的 AI 安全草案設極高門檻：五十人死亡或十億美元損失；6、英國公布數碼與科技職位計劃，十年開約二十萬個入門職位；7、全球最大主權基金淡馬錫等報告：東南亞上半年八成私人融資流向人工智能，企業採用仍落後；8、日本召開跨部門聯絡會議加快引進 AI，年內彙總行動計劃；9、南韓與日本舉行第二次人工智能工作層會議，商討安全治理與知識產權合作；10、百度搭子發布企業級智能體能力，內地智能體之爭轉向生產流程。每條均附本會顧問觀點、對日常工作的影響與一項可即時執行的建議。

本會對今日 AI 局勢的整體判斷是甚麼？

今日主線是「代理自己開門」：沃爾瑪以 MCP 網關向代理開放交易並設代理檔案與消費上限；Anthropic 發布主動式代理應用 Duet，同時換上新版 API 與 Apache 2.0 授權，把工程接線變成企業成本；百度把智能體推上生產線。另一邊，Anthropic 推銷的安全法案門檻被指過高。

這些情報的資料來源是甚麼？如何核實？

本欄每日由 AIX Society 編輯部檢索公開資料後撰稿，每條新聞均附原始來源連結（本期包括：The Verge、Unite.AI、Anthropic Release Notes（經 Releasebot 轉載）、Wain（工程分析）、The Hindu BusinessLine、IT之家（騰訊新聞轉載）、POLITICO、NPR），並將事實與本會觀點分列。讀者可按來源連結自行核實。

訂閱每週電子報

每週一封：AI 轉型的實戰觀察、工具實測與案例拆解。