2026-05-27 · OpenClaw 官方 Release
OpenClaw 2026.5.26-beta.1 把熱路徑快取、Talk 控制與可觀測性一起推進
官方 release 把可見回覆與後續慢工作拆開，重用 command／model／plugin metadata 的熱路徑快取，並讓 Talk runs 能被檢視、steer、取消或接續。很適合工程學生把互動控制、啟動路徑與 telemetry 放在同一張系統圖上。
查看來源（OpenClaw 官方 Release）→2026 年 5 月新聞檔案
本檔案保留 2026 年 5 月的精選新聞與舊聞，依日期由新到舊排列，方便學生回查每日更新脈絡。
本月每日檔案
- archives/2026-05-27.html — 2026-05-27 每日新聞檔案開啟
- archives/2026-05-26.html — 2026-05-26 每日新聞檔案開啟
- archives/2026-05-25.html — 2026-05-25 每日新聞檔案開啟
- archives/2026-05-24.html — 2026-05-24 每日新聞檔案開啟
- archives/2026-05-23.html — 2026-05-23 每日新聞檔案開啟
- archives/2026-05-22.html — 2026-05-22 每日新聞檔案開啟
- archives/2026-05-02.html — 2026-05-02 每日新聞檔案開啟
- archives/2026-05-01.html — 2026-05-01 每日新聞檔案開啟
2026-05-27 中文精選
2026-05-27 · OpenClaw 官方 Commit
OpenClaw 為 Telegram Docker E2E 補上有界 npm 安裝，提醒部署腳本也要可預測
這筆 commit 把 Telegram live／RTT Docker 腳本中的 npm 安裝流程加上邊界與對應測試，避免套件準備在教室機房、CI 或長流程中無限制拖長。課堂上可拿來比較 package preparation、timeout 設計與失敗回報。
查看來源（OpenClaw 官方 Commit）→2026-05-27 · OpenClaw 官方 Commit
OpenClaw 改用 corepack pnpm 進行 mac app 封裝，讓打包版本更可重現
mac app 封裝腳本與測試同步改成走 corepack pnpm，重點不只是「能不能打包」，而是封裝鏈版本要可釘住、可重播。對工程學生來說，這是部署工具鏈與發版一致性的直接案例。
查看來源（OpenClaw 官方 Commit）→2026-05-27 · Hermes Agent 官方 Commit
Hermes 讓 /reload-mcp 直接刷新快取中的 agent tools，不必再靠 /new 才看見新工具
gateway 重新連線 MCP 後，現在會把快取中的 AIAgent tools 與 valid_tool_names 一起更新，且保留原本的 enabled／disabled toolset 範圍。這條很適合拿來教工具探索快取、session 延續性與顯式 reload 邊界。
查看來源（Hermes Agent 官方 Commit）→2026-05-27 · TechCrunch
TechCrunch：OpenRouter 一年內估值翻倍到 13 億美元，顯示多模型路由基礎設施正被單獨定價
報導指出 OpenRouter 完成 1.13 億美元 B 輪，且六個月內使用量成長 5 倍。這對工程學生的重點不是募資本身，而是模型路由、API 正規化與多供應商切換，已經被視為獨立的 LLM 基礎設施層。
查看來源（TechCrunch）→2026-05-27 English Selected News
2026-05-27 · OpenClaw Official Release
OpenClaw 2026.5.26-beta.1 advances hot-path caching, Talk control, and observability together
The release separates visible reply delivery from slower follow-up work, reuses command/model/plugin metadata on hot paths, and makes Talk runs inspectable, steerable, cancellable, and follow-up friendly. It is a strong systems map for teaching interaction control, startup paths, and telemetry in one pass.
Source (OpenClaw Official Release) →2026-05-27 · OpenClaw Official Commit
OpenClaw bounds Telegram Docker E2E npm installs, making deployment scripts more predictable to teach
This commit adds bounded npm-install behavior and matching tests for the Telegram live/RTT Docker scripts so package preparation does not stretch unpredictably in classrooms, CI, or long-running flows. It is a clean prompt for discussing package prep, timeout policy, and failure reporting.
Source (OpenClaw Official Commit) →2026-05-27 · OpenClaw Official Commit
OpenClaw switches mac app packaging to corepack pnpm so the packaging toolchain is easier to reproduce
The mac app packaging script and its tests now go through corepack pnpm, which turns packaging from a one-off shell step into a version-pinned release pipeline. For engineering students, this is a direct lesson in packaging reproducibility and release consistency.
Source (OpenClaw Official Commit) →2026-05-27 · Hermes Agent Official Commit
Hermes makes /reload-mcp refresh cached agent tools immediately instead of requiring /new
After reconnecting MCP servers, the gateway now refreshes cached AIAgent tool lists and valid tool names while preserving each agent’s enabled/disabled toolset scope. That makes it a useful teaching case for tool-discovery caches, session continuity, and explicit reload boundaries.
Source (Hermes Agent Official Commit) →2026-05-27 · TechCrunch
TechCrunch says OpenRouter’s $1.3B valuation shows multi-model routing infrastructure is being priced as its own layer
The report says OpenRouter raised a $113M Series B and grew usage 5x in six months. The teaching value is not the funding headline alone, but the idea that routing, API normalization, and provider switching are now treated as standalone LLM infrastructure.
Source (TechCrunch) →2026-05-26 中文精選
2026-05-26 · OpenClaw 官方 Release
OpenClaw 2026.5.24-beta.2 把 thumb approval、熱路徑快取與 meeting-notes 外掛放進同一版
這版最適合拿來畫平台控制面：iMessage 可直接用 👍/👎 回覆核准，gateway 把 channel catalog、plugin metadata 與 session-store 讀取改成 process-local 快取，meeting-notes 又從核心包拆成外掛與來源契約。工程學生可把互動控制、啟動效能與資料進站一起讀。
查看來源（OpenClaw 官方 Release）→2026-05-26 · OpenClaw 官方 Commit
OpenClaw 降低 runtime metadata hot path churn，提醒學生快取與觀測訊號要一起設計
這筆 commit 直接凍結並重用 plugin metadata snapshot、package signature 與 alias map，還把 Discord realtime turn retention 延到真正開始音訊時才記錄。課堂上可追問：效能優化後，哪些 log 與 trace 仍要保留，才不會把除錯能力一起削掉。
查看來源（OpenClaw 官方 Commit）→2026-05-26 · Hermes Agent 官方 Commit
Hermes 修補 /model --global 對平面字串設定的寫回失敗，把 schema 演進變成可見的教材
當 config.yaml 把 model 寫成單一字串時，舊流程會在 gateway 靜默炸掉。這次修補先把 scalar model coercion 成 dict，再寫回 default 值並補 regression test。學生可把它當成設定格式升級、錯誤訊息可見性與持久化安全的範例。
查看來源（Hermes Agent 官方 Commit）→2026-05-26 · Hermes Agent 官方 Commit
Hermes 讓外部 secrets 每個 HERMES_HOME 只套用一次，適合拿來談啟動冪等性
load_hermes_dotenv() 過去會在多個入口重複觸發 Bitwarden 與 ASCII sanitization，讓 CLI 啟動時狀態列重複出現。新修補用 process-level guard 避免同一個 home_path 被反覆套用，剛好可拿來比較 secrets 載入、啟動噪音與長流程 refresh 的設計。
查看來源（Hermes Agent 官方 Commit）→2026-05-26 · TechCrunch
TechCrunch 用 Google 的案例提醒：AI security 是持續營運，不是發版後再補的選修題
這則外部報導適合和今天兩個專案的官方修補一起讀：快取、設定持久化、secrets、未授權資產暴露與環境變數邊界，都不是單次修正就結束。工程學生可把它轉成日常檢查表：哪些 log、allowlist、denylist 與回歸測試要固定跑。
查看來源（TechCrunch）→2026-05-26 English Selected News
2026-05-26 · OpenClaw Official Release
OpenClaw 2026.5.24-beta.2 bundles thumb approvals, hot-path caching, and a source-only meeting-notes plugin
This release is a strong systems map for class: iMessage approvals now resolve directly from 👍/👎 reactions, gateway hot paths reuse process-local metadata and session-store reads, and meeting notes move into a source-provider plugin contract. Students can read interaction control, startup performance, and ingestion as one platform story.
Source (OpenClaw Official Release) →2026-05-26 · OpenClaw Official Commit
OpenClaw reduces runtime metadata hot-path churn, making cache design and observability a paired lesson
The commit memoizes plugin metadata snapshots, package signatures, and alias maps, while deferring Discord realtime turn retention until audio actually starts. It is a good teaching case for asking how to reduce filesystem churn without erasing the diagnostics needed for debugging live systems.
Source (OpenClaw Official Commit) →2026-05-26 · Hermes Agent Official Commit
Hermes fixes /model --global persistence for flat-string configs, turning schema evolution into a visible teaching case
When config.yaml used a flat model string, the gateway path silently crashed before persisting /model --global. The fix coerces scalar model settings into a dict before mutation and adds regression coverage, making it a clean example of config-schema migration, user-facing errors, and safe persistence.
Source (Hermes Agent Official Commit) →2026-05-26 · Hermes Agent Official Commit
Hermes applies external secrets only once per HERMES_HOME, making startup idempotence easier to teach
Multiple Hermes entrypoints used to re-run external secret loading, Bitwarden fetch logic, and status printing during one startup. The new process-level guard makes secret application idempotent per HERMES_HOME, which helps students discuss startup noise, cached secret refresh, and boundary-aware initialization.
Source (Hermes Agent Official Commit) →2026-05-26 · TechCrunch
TechCrunch uses Google’s example to argue that AI security is ongoing operations, not a post-release afterthought
This external item pairs well with today’s official fixes because caching, config persistence, secrets, asset exposure, and environment-variable boundaries all need continuous hardening. It pushes students to turn news into an operations checklist: which logs, allowlists, denylists, and regression tests must run every time.
Source (TechCrunch) →2026-05-25 中文精選
2026-05-25 · OpenClaw 官方 Release
OpenClaw 2026.5.24-beta.1 把 gateway 熱路徑快取、active-run steering 與 meeting notes 管線收進同一版
這版很適合拿來畫 agent 平台的執行面：一邊是 channel catalog 與 plugin metadata 快取，另一邊是執行中的 consult 可被查詢、取消、轉向或續排，外加 meeting notes 與 Discord voice capture 的資料進站流程。工程學生可直接把效能、控制面與資料來源串成同一張系統圖。
查看來源（OpenClaw 官方 Release）→2026-05-25 · OpenClaw 官方 Commit
OpenClaw 快取 fallback provider resolution，讓多供應商降級路徑更穩定
這筆 commit 適合拿來談模型供應商切換不是只有「能不能 fallback」，還要看熱路徑查找成本、快取失效時機，以及失敗後是否能留下足夠診斷訊號。做課堂比較時，可把 provider routing、latency 與 reliability 放在同一欄分析。
查看來源（OpenClaw 官方 Commit）→2026-05-25 · Hermes Agent 官方 Release
Hermes Agent v0.14.0 仍是最新官方 release，可拿來當本地代理、工具面與工作流的基礎地圖
雖然不是今天發版，但它仍是目前最穩定的官方總覽：本地 OpenAI 相容 proxy、x_search、Teams、browser、MCP 與 cron 都在同一個工具面上。若學生要把 Hermes 當平台讀，這版仍然是理解整體能力邊界的最好入口。
查看來源（Hermes Agent 官方 Release）→2026-05-25 · Hermes Agent 官方 Commit
Hermes Agent 禁止 scratch tasks 繼承 board.default_workdir，提醒學生把工作目錄隔離當成協作安全邊界
這筆 commit 很適合拿來講多代理協作的 workspace isolation：臨時 scratch task 若直接繼承看板預設工作目錄，容易把暫時任務和正式工作區混在一起。課堂上可延伸討論 task scope、檔案寫入邊界、清理策略與回收安全。
查看來源（Hermes Agent 官方 Commit）→2026-05-25 · TechCrunch
TechCrunch 指出連 Google 也仍在即時調整 AI security，適合把新聞讀回平台防禦與營運現場
這則外部報導可當作今天的營運面補充：AI 系統安全不是部署完成後才補，而是要和 release、commit、快取、授權與隔離設計一起持續調整。學生可把它和兩個官方專案的修補節奏對照，思考哪些保護機制必須進入日常維運。
查看來源（TechCrunch）→2026-05-25 English Selected News
2026-05-25 · OpenClaw Official Release
OpenClaw 2026.5.24-beta.1 packages gateway hot-path caching, active-run steering, and meeting-notes ingestion in one release
This release is a strong systems map for class: channel-catalog and plugin-metadata caching tighten hot paths, in-flight consults can be queried, cancelled, steered, or queued, and meeting-notes plus Discord voice capture make the input pipeline visible. Students can read performance, control surfaces, and data ingestion as one platform story.
Source（OpenClaw Official Release） →2026-05-25 · OpenClaw Official Commit
OpenClaw caches fallback provider resolution, making multi-provider failover easier to reason about on the hot path
This commit is useful for teaching that fallback is not only about having a second model. Students can examine routing cost, cache invalidation timing, and whether failure paths still leave enough observability to debug degraded behavior.
Source（OpenClaw Official Commit） →2026-05-25 · Hermes Agent Official Release
Hermes Agent v0.14.0 remains the latest official release and still works as the best baseline map for local proxying, tool surfaces, and workflow plumbing
It is not a same-day release, but it is still the clearest official baseline. The local OpenAI-compatible proxy, x_search, Teams, browser, MCP, and cron features make it easier for engineering students to read Hermes as a platform with layered execution surfaces rather than as a single chat UI.
Source（Hermes Agent Official Release） →2026-05-25 · Hermes Agent Official Commit
Hermes Agent stops scratch tasks from inheriting board.default_workdir, turning workspace isolation into a concrete collaboration boundary
This commit is a clean classroom prompt for multi-agent workspace isolation. If temporary scratch work inherits the board-level default directory, transient tasks and durable workspaces blur together, so students can discuss task scope, write boundaries, cleanup behavior, and recovery safety.
Source（Hermes Agent Official Commit） →2026-05-25 · TechCrunch
TechCrunch argues that even Google is still navigating AI security in real time, which helps frame platform hardening as ongoing operations
This external report is useful as an operations-layer companion to the official repo updates. It pushes students to compare release notes and commits with the broader reality that AI security has to evolve continuously alongside caching, authorization, isolation, and recovery design.
Source（TechCrunch） →2026-05-24 中文精選
2026-05-24 · OpenClaw 官方 Release
OpenClaw 2026.5.22-beta.1 把 onboarding、Gateway 啟動、子代理上下文邊界與發佈封裝一起往前推
這版最適合拿來談平台邊界：從新手進站的啟動路徑，到 delegated worker 只繼承 AGENTS.md / TOOLS.md，再到 npm 套件瘦身與 observability smoke，都能讓學生把「可用性、委派隔離、可部署性」放在同一張系統圖上。
查看來源（OpenClaw 官方 Release）→2026-05-24 · OpenClaw 官方 Commit
OpenClaw 讓 loopback CDP 直接繞過 managed proxy，方便把本機 browser automation 的網路邊界講清楚
這筆 commit 很適合用來講瀏覽器工具鏈的實際部署問題：當控制端其實就在本機 loopback 上時，proxy 不該再插一層。課堂上可延伸比較本機除錯、遠端瀏覽器、代理設定與失敗診斷路徑。
查看來源（OpenClaw 官方 Commit）→2026-05-24 · Hermes Agent 官方 Release
Hermes Agent v0.14.0 仍是目前最新官方 release，適合當成 tools、gateway、browser、MCP 與 cron 的基礎地圖
雖然不是同日發版，但它仍是今天最穩的官方總覽。學生若要理解 agent 平台不是單一聊天介面，而是一組可擴充工具面與執行面，這版 release 仍然是最適合先讀的底圖。
查看來源（Hermes Agent 官方 Release）→2026-05-24 · Hermes Agent 官方 Commit
Hermes Agent 強化損毀 kanban.db 備份流程，直接把 path injection 與資料恢復邊界變成教學案例
這筆 commit 很適合提醒學生：多代理協作板一旦落地成資料庫，修復策略本身也會變成安全面。可拿來討論損毀偵測、備份路徑驗證、CodeQL 發現的輸入風險，以及「恢復」不能比「損毀」更危險。
查看來源（Hermes Agent 官方 Commit）→2026-05-24 · InfoQ
InfoQ 報導 xAI 釋出 Grok Skills，並更新 Tool Calling Responses API
這則外部消息很適合補上平台 API 視角：當技能與 tool-calling 變成官方產品表面，工程學生就能更具體比較「模型能力」和「可被外部系統穩定呼叫的代理能力」之間的差異。
查看來源（InfoQ）→2026-05-24 English Selected News
2026-05-24 · OpenClaw Official Release
OpenClaw 2026.5.22-beta.1 advances onboarding, Gateway startup paths, delegated-worker context boundaries, and release packaging together
This release is useful as a boundary-mapping exercise. Students can trace how first-run setup, gateway startup, subagent bootstrap isolation, package size, and observability checks all shape whether an agent platform is teachable, operable, and safe to extend.
Source（OpenClaw Official Release） →2026-05-24 · OpenClaw Official Commit
OpenClaw bypasses the managed proxy for loopback CDP, clarifying the network boundary for local browser automation
This commit is a clean classroom example of infrastructure fit. If the browser control endpoint is already local loopback, an extra proxy layer can complicate debugging and failure modes, so students can compare local automation, remote browsers, proxying, and diagnostics.
Source（OpenClaw Official Commit） →2026-05-24 · Hermes Agent Official Release
Hermes Agent v0.14.0 remains the latest official release and still works as the best baseline map for tools, gateway behavior, browser, MCP, and cron
It is not a same-day release, but it remains the clearest official foundation snapshot. For engineering students, it is still the best starting point for reading Hermes as an extensible agent platform rather than as a single chat interface.
Source（Hermes Agent Official Release） →2026-05-24 · Hermes Agent Official Commit
Hermes Agent hardens corrupt kanban.db backup handling, turning path injection and recovery boundaries into a concrete systems lesson
This commit is a good prompt for discussing persistence safety in multi-agent coordination. Once task state lives in a database, corruption recovery has to defend its own file paths and trust boundaries instead of assuming backup logic is automatically safe.
Source（Hermes Agent Official Commit） →2026-05-24 · InfoQ
InfoQ reports that xAI released Grok Skills and updated its Tool Calling Responses API
This external item helps students compare model capability with agent surface design. Once skills and tool-calling become explicit product primitives, the engineering discussion can move from “what the model can do” to “what outside systems can invoke reliably.”
Source（InfoQ） →2026-05-23 中文精選
2026-05-23 · OpenClaw 官方 Release
OpenClaw 2026.5.20 把執行核准、Discord 語音跟隨、Policy 外掛與多供應商治理收進同一版
這條官方 release 適合拿來畫 agent 平台總覽：同一版同時碰到 approvals、realtime voice、policy 檢查與 provider 路由，學生可直接拆成互動層、治理層與模型供應商層三條線。
查看來源（OpenClaw 官方 Release）→2026-05-23 · OpenClaw 官方 Commit
OpenClaw 改用檔案系統安全的 trash 流程刪除 agent，讓「刪除」變成可討論的資料生命週期問題
這筆 commit 很適合提醒學生：agent 系統不只要會建立狀態，也要安全地移除狀態。討論時可把刪除流程拆成 recoverability、權限邊界、誤刪風險與稽核軌跡。
查看來源（OpenClaw 官方 Commit）→2026-05-23 · Hermes Agent 官方 Release
Hermes Agent v0.14.0 把可攜安裝、本地 OpenAI 相容代理、多工具面與平台穩定性整理成基礎版面
雖然不是今天發版，但它仍是目前最新的官方 release，適合先當成結構底圖：從安裝與 gateway，到 browser、MCP、Google Workspace、kanban 與 cron，已經是可拿來做課堂實作的 agent 基礎設施。
查看來源（Hermes Agent 官方 Release）→2026-05-23 · Hermes Agent 官方 Commit
Hermes Agent 為 MiniMax OAuth 改成每次請求都刷新短效 access token，讓供應商授權生命週期更可控
這筆 commit 很適合拿來談 provider 整合的真實摩擦：短效 token 若不跟著請求節奏刷新，agent 可能在長流程中途失效。課堂上可延伸比較 refresh 時機、錯誤回報與降級策略。
查看來源（Hermes Agent 官方 Commit）→2026-05-23 · InfoQ
InfoQ 整理 Cloudflare 六層 agent infrastructure stack，適合對照 compute、orchestration、memory 與 browsing
這則外部報導很適合補平台視角：Cloudflare 把 compute、workflow、memory、browser run 與 commerce 組成同一棧，學生可直接拿來對照 OpenClaw 與 Hermes 各自覆蓋到哪些層。
查看來源（InfoQ）→2026-05-23 English Selected News
2026-05-23 · OpenClaw Official Release
OpenClaw 2026.5.20 packages exec approvals, Discord voice-follow behavior, a policy plugin, and broader provider governance in one release
This release is the clearest official systems map for class today. One version touches approval flow, realtime voice sessions, policy-backed channel checks, and provider routing, so students can separate interaction surfaces, governance controls, and model-plumbing layers.
Source (OpenClaw Official Release) →2026-05-23 · OpenClaw Official Commit
OpenClaw switches agent deletion to an fs-safe trash path, turning “delete” into a data-lifecycle topic
This commit is a good prompt for teaching that agent systems must remove state safely as well as create it. Students can break the change into recoverability, permission boundaries, accidental-deletion risk, and auditability.
Source (OpenClaw Official Commit) →2026-05-23 · Hermes Agent Official Release
Hermes Agent v0.14.0 packages portable install, a local OpenAI-compatible proxy, broader tool surfaces, and platform reliability into a foundation release
It is still the latest official release and works well as the baseline map for students. From install and gateway behavior to browser, MCP, Google Workspace, kanban, and cron, the release reads like an agent infrastructure stack instead of a single-feature drop.
Source (Hermes Agent Official Release) →2026-05-23 · Hermes Agent Official Commit
Hermes Agent refreshes short-lived MiniMax OAuth access tokens per request, making provider auth lifecycle easier to reason about
This commit is a strong classroom example of the friction inside provider integrations. If short-lived tokens are not refreshed at the right boundary, long-running agent flows fail mid-task, so students can discuss refresh timing, error reporting, and degraded behavior.
Source (Hermes Agent Official Commit) →2026-05-23 · InfoQ
InfoQ reports Cloudflare’s six-layer agent infrastructure stack, giving students a clean compare-and-contrast map for compute, orchestration, memory, and browsing
This external item adds the platform-comparison angle. Cloudflare’s stack groups compute, workflows, memory, browser execution, and commerce in one model, which makes it easier for students to compare how OpenClaw and Hermes cover different infrastructure layers.
Source (InfoQ) →2026-05-22 中文精選
2026-05-22 · OpenClaw 官方 Release
OpenClaw 2026.5.20 把執行核准、Discord 語音跟隨、Policy 外掛與多供應商治理一起推進到同一版
這版很適合當成今天的系統總覽案例：同一份 release 同時碰到 exec approvals、即時語音 session、policy-backed channel 檢查，以及 xAI / OpenRouter 等 provider 路由。對工程學生而言，重點不是功能清單，而是看 agent 平台如何把互動、治理與供應商整合放進同一條發布節奏。
查看來源（OpenClaw 官方 Release）→2026-05-22 · OpenClaw 官方 Commit
OpenClaw 強化 Codex sandbox execution，提醒學生把 coding agent 放進受限執行邊界來看
這筆 commit 適合拿來談 coding agent 的安全面：當 agent 能直接碰 shell、檔案與外部工具時，真正的工程問題會變成 sandbox 規則、權限縮限、失敗回報與可追蹤性。做課堂比較時，可把「模型會不會寫」和「系統敢不敢放它執行」分開分析。
查看來源（OpenClaw 官方 Commit）→2026-05-22 · Hermes Agent 官方 Release
Hermes Agent v0.14.0 把可攜安裝、多模型路由、工具面擴充與平台穩定性整理成基礎版面
雖然不是今天發布，但它仍是目前最新的官方 release，適合工程學生先建立整體地圖：從安裝與 gateway，到 browser / MCP / Google Workspace / kanban / cron，全都已經不是單點功能，而是可延伸成課堂實作的基礎設施。
查看來源（Hermes Agent 官方 Release）→2026-05-22 · Hermes Agent 官方 Commit
Hermes Agent 新增 Bitwarden Secrets Manager 整合，把祕密管理正式拉進 agent 工作流
這筆 commit 很適合補上「agent 接到真實工作後，憑證到底放哪裡」這個常被忽略的問題。當 secrets manager 進入工具鏈，學生就能具體討論 API key 載入、最小權限、輪替、稽核與失敗降級，而不是把敏感資訊繼續硬寫在設定檔。
查看來源（Hermes Agent 官方 Commit）→2026-05-22 · The Verge
The Verge 報導 Spotify Studio 用 AI agent 幫使用者組每日 podcast 與 briefing
這則外部消息適合補上「連接個人資料來源後，agent 如何變成產品」：一旦 agent 要讀信箱、行事曆或聆聽偏好來生成 briefing，工程重點就會轉向資料接線、個人化邏輯、輸出驗證與隱私邊界，而不只是一段 prompt。
查看來源（The Verge）→2026-05-22 English Selected News
2026-05-22 · OpenClaw Official Release
OpenClaw 2026.5.20 ships exec approvals, Discord voice-follow behavior, a policy plugin, and broader provider governance in one release
This is the best official overview item for class today because one release spans approval flow, realtime voice sessions, policy-backed channel checks, and provider routing. Students can read it as a systems map showing how interaction surfaces, governance, and model plumbing evolve together in an agent platform.
Source（OpenClaw Official Release） →2026-05-22 · OpenClaw Official Commit
OpenClaw hardens Codex sandbox execution, making the execution boundary itself a teaching topic
This commit is useful for teaching the security side of coding agents. Once an agent can touch shell commands, files, and external tools, the engineering problem becomes sandbox rules, permission narrowing, failure handling, and auditability rather than “can the model write code” alone.
Source（OpenClaw Official Commit） →2026-05-22 · Hermes Agent Official Release
Hermes Agent v0.14.0 packages portable install, model routing, tool-surface expansion, and platform reliability into a foundation release
It is not a same-day release, but it remains the latest official release and still works well as the structural baseline for students. From install and gateway behavior to browser, MCP, Google Workspace, kanban, and cron, the release reads like an agent infrastructure stack rather than a single-feature drop.
Source（Hermes Agent Official Release） →2026-05-22 · Hermes Agent Official Commit
Hermes Agent adds Bitwarden Secrets Manager integration, bringing secret handling into the workflow conversation
This commit is a strong classroom prompt for asking where credentials belong once agents move into real work. With a secrets manager in the toolchain, students can discuss API-key loading, least privilege, rotation, audit trails, and degraded behavior instead of treating secrets as config-file clutter.
Source（Hermes Agent Official Commit） →2026-05-22 · The Verge
The Verge reports that Spotify Studio uses an AI agent to assemble daily podcasts and briefings for users
This external item adds the productization layer. Once an agent reads inboxes, calendars, or listening history to generate a personal briefing, the engineering discussion has to include data connectors, personalization logic, output verification, and privacy boundaries rather than stopping at prompt quality.
Source（The Verge） →2026-05-02 中文精選
2026-05-02 · OpenClaw 官方 Release
OpenClaw 2026.4.29 把主動 steering、people-aware memory、NVIDIA provider 與通道穩定性整合成同一波系統更新
這版仍是今天最值得當作總覽案例的官方來源：同一份 release 同時牽動互動控制、記憶、provider 與 channel，適合學生先畫出 agent 系統圖，再判斷哪些功能是面向使用者、哪些其實是支撐可靠性的底層工程。
查看來源（OpenClaw 官方 Release）→2026-05-02 · OpenClaw 官方 Commit
OpenClaw 把 message access groups 一般化，讓通道受眾與訊息可見範圍更容易拆開管理
這筆 commit 適合拿來教「agent 不只是會回覆，還要管誰看得到什麼」。當 access groups 抽象化之後，Discord DM、頻道 audience 與後續通道擴充會更容易共用同一套權限模型，學生可順便討論 ACL、租戶隔離與稽核需求。
查看來源（OpenClaw 官方 Commit）→2026-05-02 · Hermes Agent 官方 Release
Hermes Agent v0.12.0 把 Autonomous Curator、自我改進回路與更多 provider / skill 整理成正式主線版本
對工程學生來說，這條官方 release 很適合拿來討論「agent 平台開始維護自己的工具鏈」意味著什麼：技能庫評分、整理與 provider 擴充已不是附加功能，而是平台治理與可維運性的主線能力。
查看來源（Hermes Agent 官方 Release）→2026-05-02 · Hermes Agent 官方 Commit
Hermes Agent 為 Curator 補上延後首次執行與 --dry-run 預覽，讓技能整理先可觀察再自動化
這筆 commit 很適合帶進課堂，因為它把「先看再放手給 agent 做」寫進產品流程。當自動整理技能庫之前先提供 dry-run 預覽，學生就能更具體討論 rollout、人工覆核、可回滾性與背景工作排程。
查看來源（Hermes Agent 官方 Commit）→2026-05-02 · The Verge
The Verge 報導 Microsoft 想讓律師在 Word 文件裡信任新的 AI agent
這則外部消息適合補上「領域型 agent 要怎麼取得信任」：一旦 agent 直接進入文件審閱流程，工程重點就不只剩模型能力，還要把引用來源、修改可追蹤性、審批流程與責任界線一起設計進去。
查看來源（The Verge）→2026-05-02 English Selected News
2026-05-02 · OpenClaw Official Release
OpenClaw 2026.4.29 packages active steering, people-aware memory, NVIDIA onboarding, and channel reliability into one system release
This release remains the best official overview item for class because it moves interaction control, memory, provider coverage, and channel reliability together. Students can start by mapping user-facing behavior against the infrastructure layers that keep the agent usable.
Source（OpenClaw Official Release） →2026-05-02 · OpenClaw Official Commit
OpenClaw generalizes message access groups so channel audiences and message visibility can be managed as separate controls
This commit is useful for teaching that an agent system is not only about generating replies but also about deciding who can see which messages. Once access groups become a reusable abstraction, Discord DMs, channel audiences, and future integrations can share a cleaner authorization model.
Source（OpenClaw Official Commit） →2026-05-02 · Hermes Agent Official Release
Hermes Agent v0.12.0 formalizes the Autonomous Curator, a stronger self-improvement loop, and broader provider / skill coverage
For engineering students, this release is a compact case study in what it means when an agent platform starts maintaining parts of its own toolchain. Skill curation, self-improvement, and provider expansion are no longer side features but core governance and operability concerns.
Source（Hermes Agent Official Release） →2026-05-02 · Hermes Agent Official Commit
Hermes Agent adds deferred first-run behavior and a --dry-run preview to Curator so skill maintenance can be observed before it is automated
This commit is a strong classroom prompt because it bakes “observe first, automate second” into the product flow. A dry-run preview before autonomous skill cleanup makes it easier to discuss rollout safety, human review, reversibility, and background scheduling.
Source（Hermes Agent Official Commit） →2026-05-02 · The Verge
The Verge reports that Microsoft wants lawyers to trust a new AI agent inside Word documents
This external item adds the trust boundary for domain-specific agents. Once an agent participates directly in document review, the engineering discussion has to include source attribution, change traceability, approval flow, and accountability instead of stopping at raw model capability.
Source（The Verge） →2026-05-01 中文精選
2026-05-01 · OpenClaw 官方 Release
OpenClaw 2026.4.29 把主動 steering、people-aware memory、NVIDIA provider 與通道穩定性一起推進
這版很適合工程學生拆成四層：訊息佇列開始預設支援 active-run steering，memory 不再只記片段而是朝可追溯人物知識走，NVIDIA 與 manifest-backed model/auth 路徑擴大供應商覆蓋，Slack／Telegram／Discord 等通道則持續補齊可靠性。看 release 時可把「互動控制、記憶、provider、channel」放在同一張系統圖上。
查看來源（OpenClaw 官方 Release）→2026-05-01 · OpenClaw 官方 Commit
OpenClaw 修補 bundled runtime deps 的重裝迴圈，避免 gateway 預熱後又被 lazy plugin activation 重新觸發 pnpm install
這筆 commit 很適合拿來教 plugin 供應鏈穩定性：如果系統已經有較完整的 package-level runtime deps，後續啟用外掛時卻又誤判成未收斂並重裝依賴，就會拖慢啟動、污染快取，也讓教室機房或 CI runner 的行為變得不可預測。
查看來源（OpenClaw 官方 Commit）→2026-05-01 · Hermes Agent 官方 Release
Hermes Agent v0.12.0 把 Autonomous Curator、升級版 self-improvement loop 與更多 provider / skill 變成正式主線能力
這次 release 最值得課堂討論的是「agent 開始維護自己的 agent 工具鏈」：Curator 會週期性評分與整理技能庫，self-improvement loop 改成更偏 rubric 與 active-update，LM Studio 也升格為一級 provider。對工程學生而言，這是觀察自動化維護、工具治理與多供應商推論整合的好案例。
查看來源（Hermes Agent 官方 Release）→2026-05-01 · Hermes Agent 官方 Commit
Hermes Agent 新增 durable multi-profile collaboration board，把 kanban 任務板、dispatcher、worker skills 與 dashboard 串成可持久化協作面
這筆 commit 很適合教多代理協作不是只靠 delegate，而是需要可落盤的任務狀態、CAS claim、heartbeat、run history 與可觀測的 UI。當 task、link、comment、dispatcher 與 dashboard 都進入同一套資料結構，學生就能更具體比較「聊天式委派」和「流程式協作」的差異。
查看來源（Hermes Agent 官方 Commit）→2026-05-01 · TechCrunch
TechCrunch 報導 Stripe 推出可供 autonomous AI agents 使用的 Link 數位錢包
這則外部消息很適合補上交易與權限邊界：當 agent 不只是找資料或呼叫工具，而是能碰到付款與購物流程，系統設計就得把身份確認、授權範圍、風險控管、交易回報與人工覆核一起納入。做專題時可把「工具可用」和「可安全成交」分開討論。
查看來源（TechCrunch）→2026-05-01 English Selected News
2026-05-01 · OpenClaw Official Release
OpenClaw 2026.4.29 advances active steering, people-aware memory, NVIDIA onboarding, and channel reliability in one release
This release is a strong teaching case because four system layers move together: active-run message steering becomes a first-class interaction control, memory grows toward a provenance-aware people wiki, NVIDIA and manifest-backed model/auth flows widen provider coverage, and Slack / Telegram / Discord fixes keep the operational surfaces usable. Students can map the release as interaction control, memory, provider, and channel engineering rather than as a single feature drop.
Source（OpenClaw Official Release） →2026-05-01 · OpenClaw Official Commit
OpenClaw stops bundled runtime dependency reinstall loops after gateway pre-staging and later lazy plugin activation
This commit is useful for teaching plugin supply-chain stability: if the platform already has a materialized package-level superset of runtime dependencies, later lazy activation should not mis-diagnose the state and relaunch pnpm install. Otherwise startup gets slower, caches drift, and classroom or CI environments stop behaving reproducibly.
Source（OpenClaw Official Commit） →2026-05-01 · Hermes Agent Official Release
Hermes Agent v0.12.0 lands the Autonomous Curator, a stronger self-improvement loop, and broader provider / skill coverage
The main classroom lesson is that an agent platform is starting to maintain parts of its own toolchain. Curator now grades and consolidates skills in the background, the self-improvement loop becomes more rubric-driven and active-update biased, and LM Studio plus other providers become more first-class. It is a compact case for discussing automation governance, tool maintenance, and multi-provider systems design together.
Source（Hermes Agent Official Release） →2026-05-01 · Hermes Agent Official Commit
Hermes Agent adds a durable multi-profile collaboration board with kanban storage, dispatcher, worker skills, and dashboard UI
This commit is a good teaching prompt for multi-agent coordination beyond simple delegation. Once tasks, links, comments, claims, heartbeats, run history, and dashboard updates share the same persisted substrate, students can compare “chat-driven delegation” with “workflow-driven collaboration” in a much more concrete way.
Source（Hermes Agent Official Commit） →2026-05-01 · TechCrunch
TechCrunch reports that Stripe introduced a Link digital wallet that autonomous AI agents can use
This external report adds the transaction and permission boundary. When agents move from retrieval and tool calls into payment-capable workflows, engineering analysis has to include identity checks, authorization scope, fraud control, transaction feedback, and human review instead of discussing “agent capability” as a single bucket.
Source（TechCrunch） →