發布日期：2026-07-01
以前你派子代理（subagent，幫你跑腿的 AI 小助手）去工作時，你必須等它做完才能繼續跟主代理講話，現在它們預設全都在背景獨立運作了！主代理可以繼續幫你做別的事，等小助手做完再來回報。而且，如果小助手是在隔離的工作環境（worktree）裡改程式碼，它現在做完會自動幫你 commit、push 甚至開好草稿 PR（pull request，請求合併程式碼的流程），不用再停下來等你點頭。這就像你原本只能盯著助理把資料打完才能派下一個工作，現在你可以一口氣派十個助理各自回座位趕工，他們做完還會自動把報告放在你的桌上，不浪費你一點時間。
現在你可以在更多地方順暢使用 Claude 了！大家期待已久的「Claude in Chrome」功能正式上線，讓你可以直接在瀏覽器裡無縫操作。如果你有接各種不同的伺服器，現在系統加入了 AWS 上的 Claude Platform（企業雲端通道），如果在第一順位的伺服器找不到你要的模型，它現在聰明到會自動往下一個備用伺服器尋找（failover chain），不會直接兩手一攤說找不到。此外，用 AWS 或 Mantle 登入時，如果金鑰過期，它也會自動幫你重新刷新（awsAuthRefresh），不會再跑到一半突然卡死叫你重新登入。這就像你拿著提款卡，以前這台機器沒錢它就叫你回家，現在它會自動幫你連到隔壁銀行的金庫，而且卡片到期還會自動幫你展延。
這次把很多視覺上的小毛病都修好了。包含 Markdown 的表格在全螢幕下終於不會超出邊界、程式碼的顏色高亮也升級得更精準。如果你用鍵盤連點網址，現在會連 https:// 整個一起選起來，Mac 用戶用 Cmd+點擊 也能正確在全螢幕下打開網頁了。在任務面板上，背景任務做完後不會再一直卡在「執行中」騙你；如果你切換了 Git 分支，修改差異（/diff）的面板也會乖乖跟著更新，不會死當在那邊。就連你在 API 重試時，以前只會看到無盡的轉圈圈，現在如果伺服器大塞車，它會直接給你一個狀態網頁的連結。這就像戴上度數正確的眼鏡，以前看表格會變形、顏色會糊掉，現在一切都清晰明瞭，連路標指引都寫得清清楚楚。
這次改善了網路與代理人連線的穩定度。以前如果在 AI 回答到一半時網路稍微閃斷，整個任務就會因為 ECONNRESET 等暫時性錯誤而直接失敗；現在它懂得稍等一下再自動重試（backoff），不會直接放棄。對於在沙盒（sandbox，隔離的安全執行環境）裡重複存取同一個網站的程式，它也不會再瘋狂發送背景分類請求浪費資源。在 Mac 上跑背景代理人時，原本常會遇到連不上區域網路主機，或是只要打開面板就會每隔 52 秒跳出「重新連線」的毛病，現在都透過修復權限與機制解決了。如果你讓多個代理人組隊工作，以前遇到 API 錯誤死掉的隊友只會默默失蹤，現在它會向隊長回報「失敗」，而且你只要傳個訊息給卡住的隊友，就能立刻喚醒它重新嘗試。就像以前打電話遇到訊號不好就直接斷線，或是同事出包只會裝死；現在它不只會自動重撥，遇到問題的團隊成員也會主動舉手求救，等你拍拍肩膀就繼續上工。
這次修好了一堆讓你在切換指令和視窗時卡住的小毛病。比如以前你在計畫模式（plan mode，讓 AI 先寫計畫再執行的模式）下啟動，它不會自動放行唯讀檔案的工具呼叫權限；或者當你用 /branch 建立對話分支時，它會用壓縮過的摘要來當作預設名稱而不是你最初的提示，現在這些都修正了。如果你在終端機裡用 claude attach 連線進代理人畫面，按左鍵退回不會再錯誤地把你踢回終端機；使用 /desktop 或是透過捷徑（symlink）路徑讀取 .claude/rules/，也不會再發生找不到工作目錄或載入失敗的靈異事件。同時，你現在可以直接從代理人管理介面點 /login 登入，而且在 SDK 與桌面版查看工作流進度時，最先出發的任務也不會再莫名其妙從清單上消失，最後，也阻止了 claude --bg 跟 --print 一起使用時產生無法連線的幽靈任務。就像開車導航，以前稍微偏離路線或切換視窗就會當機找不到路，甚至按個返回鍵就把你踢下車，現在儀表板顯示準確，想去哪都能順暢抵達。
想設計圖表或儀表板時，現在可用 /dataviz 取得指引與配色檢查。它甚至附帶了一個可以直接跑的色系驗證器，確保你選出來的顏色搭配不會糊成一團。此外，舊版的 /agents 設定精靈（wizard，一步步引導你設定的工具）已經功成身退被移除了；以後想建立或管理代理人，直接開口叫 Claude 幫你處理，或是自己去編輯 .claude/agents/ 資料夾裡的檔案就好。就像公司以前配發的繁瑣紙本申請單被廢除了，現在你要什麼圖表或想聘請新助理，直接跟秘書交代一聲，她就會幫你辦到好，而且還附贈一位超強的圖表配色設計師。
When you dispatch a subagent (an AI assistant running errands for you), you no longer have to twiddle your thumbs waiting for it to finish before talking to the main agent again; they now run independently in the background by default! The main agent can keep helping you while the assistants work and report back when done. Moreover, if an assistant modifies code inside an isolated Git worktree, it will now automatically commit, push, and open a draft PR (pull request, the process for merging code) when it finishes, instead of pausing to ask for your permission. It's like upgrading from standing over your assistant's shoulder while they type, to dispatching ten assistants to their desks simultaneously; they do the work in parallel and automatically leave the finished reports on your desk without wasting a second of your time.
You can now use Claude smoothly across even more platforms! The highly anticipated "Claude in Chrome" feature is officially generally available, letting you operate seamlessly right in your browser. If you use multiple server setups, the system now includes Claude Platform on AWS (an enterprise cloud channel) as a provider. Even better, if the primary server doesn't have the AI model you requested, it's now smart enough to automatically try the next backup server in line (failover chain) rather than just giving up. Furthermore, when logging in via AWS or Mantle, if your authentication key expires mid-task, it will automatically refresh it (awsAuthRefresh) instead of halting the program and demanding you log in again. It's like using an ATM: previously, if a machine was empty, it sent you home; now it automatically wires the cash from a neighboring bank's vault, and gracefully renews your card if it happens to expire while you're standing there.
A slew of visual glitches and UI dead-ends have been resolved. Markdown tables will finally wrap cleanly instead of overflowing the screen in fullscreen mode, and syntax highlighting for code has been noticeably sharpened. If you double-click a URL, it will now select the entire address including the https:// part, and Mac users can successfully Cmd-click links to open them even when in fullscreen mode. On the dashboard side, background tasks will no longer endlessly deceive you with a stuck "Running" status after they actually finish; and if you switch Git branches outside the app, your /diff panel will obediently refresh instead of locking up with stale data. Even the API retry experience has improved: instead of a never-ending spinning wheel during a server traffic jam, it will explicitly hand you a link to the server status page. It's like putting on a proper pair of glasses: tables and colors that used to distort or blur are now crisp and readable, and broken street signs have been replaced with clear, helpful directions.
We've improved network and agent connection stability. Previously, a brief network drop mid-response (like an ECONNRESET transient error) would instantly abort the entire turn; now it knows to back off and auto-retry instead of giving up. For sandboxed processes (isolated secure execution environments) repeatedly accessing the same host, it no longer spams excessive background classifier requests. On Macs running background agents, the frustrating "no route to host" for local networks and the "Reconnecting…" message flashing every 52 seconds when the panel was open have been fixed. If you run agent teams, a teammate that dies from an API error will now explicitly report "failed" to the lead instead of disappearing, and messaging a stuck teammate will instantly wake it up to retry. It's like a phone call that used to hang up at the slightest static, or a coworker who vanished when stuck; now the call auto-redials, and the coworker raises their hand for help, ready to resume as soon as you tap their shoulder.
We've patched a bunch of minor glitches that caused hangups when switching commands and views. For example, starting a session in plan mode (where the AI writes a plan before executing) will now properly auto-allow read-only tool calls; and when you create a conversation branch using /branch, it now correctly derives its default name from your first real prompt instead of the compacted summary. If you use claude attach <id> to join an agent view, pressing the left arrow no longer falsely kicks you all the way back to the shell. The mysterious errors where /desktop couldn't determine the working directory after worktree changes, or where .claude/rules/ failed to load via symlinked paths, are fixed. Furthermore, you can now trigger /login directly from the agents view, the earliest agents no longer drop off the workflow progress list in SDK/desktop apps, and running claude --bg with --print is now correctly rejected upfront instead of silently creating an unattachable ghost session. It's like using a car's GPS: it used to freeze if you took a slight detour or kicked you out of the car if you pressed 'back', but now the dashboard is fully reliable and takes you exactly where you want to go.
When designing charts or dashboards, you can now use /dataviz for guidance and color checking. It comes complete with a runnable color-palette validator to ensure your color choices don't blend into an unreadable mess. Meanwhile, the old /agents wizard (the step-by-step setup tool) has been retired. From now on, if you want to create or manage subagents, simply ask Claude to do it for you, or edit the files in the .claude/agents/ directory directly. It's like your company finally abolishing those tedious multi-page paper request forms; now, whether you need a fresh chart or want to hire a new assistant, you just tell your secretary what you want and it's done—and you get an expert color designer thrown in for free.
Subagents now run in the background by default, so Claude keeps working while they run and is notified when they finish (previously a gradual rollout)
Added background agent notifications in `claude agents` — sessions that need input or finish now fire the `Notification` hook (`agent_needs_input` / `agent_completed`)
Background agents launched from `claude agents` now commit, push, and open a draft PR when they finish code work in a worktree, instead of stopping to ask
The built-in Explore agent now inherits the main session's model (capped at opus) instead of running on haiku
Subagents and context compaction now inherit the session's extended thinking configuration, improving output quality on delegated tasks
Subagents now treat messages from the agent that launched them as normal task direction; an agent's message is still never treated as the user's approval
Improved focus mode: subagents launched in a turn now appear in its activity summary, and completed background notifications fold into a single countClaude in Chrome is now generally available
Gateway: added Claude Platform on AWS (anthropicAws) as an upstream provider; model-not-found responses now advance the failover chain
Fixed Claude Platform on AWS and Mantle sessions dead-ending with "Please run /login" when the STS token expires — `awsAuthRefresh` now runs automaticallyFixed background tasks in web, desktop, and VS Code task panels getting stuck on "Running" after they finish or after resuming a session
Fixed the `/diff` panel not refreshing when you switch branches or commit outside the session
Fixed markdown tables overflowing and wrapping their right border when rendered in fullscreen mode
Fixed Cmd+click not opening URLs in fullscreen mode in Warp on macOS
Fixed double-click word selection in fullscreen mode to select the entire URL including the scheme
Improved syntax highlighting accuracy in code blocks, diffs, and file previews by upgrading to highlight.js 11
Improved API retry UX: the error reason is now shown after the second attempt, and a status page link replaces the spinner tip when the API is overloaded
Keyboard shortcut hints now show opt/cmd instead of alt/super when connected from a Mac over SSHFixed brief network drops mid-response aborting the turn — transient errors like ECONNRESET now retry with backoff instead of failing
Fixed excessive background classifier requests when sandboxed processes repeatedly accessed the same network host
Fixed "no route to host" for local-network hosts in macOS background agent sessions by declaring Local Network entitlements
Fixed background agents repeatedly showing "Reconnecting…" every ~52 seconds on macOS while the agents view was open
Fixed agent teams: a teammate that dies on an API error now reports "failed" to the lead, and messaging a stuck teammate wakes it to retry immediatelyFixed plan mode not auto-allowing read-only tool calls when a session starts in plan mode
Fixed `/branch` deriving its default fork name from the compaction summary instead of the first real prompt
Fixed pressing `←` inside `claude attach <id>` exiting to the shell instead of opening the agent view
Fixed `/desktop` failing with "Cannot determine working directory" after entering and exiting a worktree
Fixed `.claude/rules/` conditional rules not loading when the target file is reached via a symlinked path
`/login` now opens the sign-in dialog from the `claude agents` view instead of saying it isn't available
Fixed the workflow progress view dropping the earliest agents from the list while the phase counter stayed correct in SDK and desktop-app sessions
Fixed `claude --bg` silently creating an unattachable session when combined with `--print`/`-p`; the conflicting flags are now rejected up frontAdded `/dataviz` skill for chart and dashboard design guidance with a runnable color-palette validator
Removed the `/agents` wizard; ask Claude to create or manage subagents, or edit `.claude/agents/` directly