Anthropic Upgrades Claude Code With Autonomous Mode and Git Worktrees

Anthropic升级Claude Code：引入自主模式与Git Worktrees 新增的安全后台自动化、项目专属记忆及远程工作流，使该工具无需开发者人工干预即可独立处理长时间运行的开发任务。 本次演示概述了Claude Code围绕开发者体验与自主性进行的重大升级，涵盖移动端远程控制链接、无闪烁的虚拟化终端UI以及交互式桌面GUI等改进。更新还引入了多项高级自动化功能，包括用于独立安全分类的Auto Mode、用于隔离并行开发的Git Worktrees，以及用于维持持久项目上下文的Auto Memory。此外，视频还展示了多智能体代码审查，以及由本地循环和在Anthropic服务器上独立执行的远程例程驱动的自动化工作流。 Claude Code 推出“远程控制”（Remote Control）功能，开发者可通过生成的 Web 链接在移动端实时监控和交互终端任务，并引入基于虚拟化滚动的全屏模式以彻底消除终端闪烁。 (0:00) 全新改版的 Desktop GUI 引入分屏拖拽与多会话管理，并提供 Plan、Diff 与 Files 专用交互视图，以及可自动生成会话索引目录的“固定为章节”（pin as chapter）实验功能。 (5:00) “自动模式”（Auto Mode）利用内部安全分类器自动批准安全的工具执行，同时优化的 Git Worktrees 机制允许共享 node_modules 目录，实现多 Agent 并行独立开发。 (10:00) 借助 Auto Memory 功能，Claude 可通过维护 memory.md 索引持久化存储项目特定规范与上下文，并配合多智能体审查流程执行更深入的代码评审。 (15:00) 新增本地 /loop 命令与运行在云端的 Routines 自动化功能，支持通过定时调度或 Webhook 在后台全天候执行自主任务，无需保持本地设备开机。 (20:00)

Anthropic technical staff member Dixon Tsai outlines updates to Claude Code focused on two main priorities: refining the developer experience through responsive terminal and desktop interfaces, and expanding agent autonomy with isolated workspaces, automatic permissions, and background routines. Flicker-Free Terminal and Cross-Device Mobility (0:00) To address common interface bottlenecks, Claude Code introduces a fullscreen mode that virtualizes terminal scrollback. This prevents output flickering and misalignment caused by appending continuous text, preserves clickable interface elements, and maintains flat memory consumption during long sessions. A voice mode, activated via the `/voice` command and held spacebar, enables direct dictation. For distributed and continuous workflows, the new 'remote control' feature allows developers to transition active sessions from desktop terminals to web browsers or mobile devices. Initiated via a terminal command that generates a web link, this mechanism synchronizes live terminal state and allows users to dispatch subagents and monitor long-running, multi-pull-request sessions remotely. Desktop GUI and Structured Context Management (5:00) The updated desktop interface targets multi-task organization with a redesigned sidebar featuring project grouping, filtering, and a drag-and-drop split view for simultaneous sessions. To improve visibility during complex tasks, dedicated views in the interface separate execution into distinct stages: a 'plan' view for marking steps and leaving notes, a 'diff' view for line-by-line code review, and a 'files' view for manual directory edits. To facilitate navigation in extensive sessions, desktop lead Anthony Morris introduced an experimental 'pin as chapter' feature. This tool allows users to generate semantic headings from assistant messages to create a dynamic table of contents across lengthy transcripts. Autonomy via Auto Mode and Git Worktrees (10:00) To prevent autonomous tasks from stalling on permission checks, Anthropic implemented 'Auto Mode.' This setting uses an internal classifier to evaluate tool calls, automatically approving safe operations while halting only for destructive actions or potential prompt injections. To allow multiple agents to work concurrently without conflict, Claude Code integrates Git worktrees. The implementation mitigates traditional storage inefficiencies by letting developers declare shared directories, such as `node_modules`. Claude can manage feature branches autonomously through built-in tools that open an isolated worktree, execute changes, and clean up the environment upon completion. Persistent Memory, Multi-Agent Review, and Background Automation (15:00) Claude Code addresses repetitive context setting and token expenditure via 'Auto Memory.' Relying on reasoning models like Claude 3.7 Sonnet, the system maintains a persistent `memory.md` directory index that accumulates build commands, debugging notes, and repository conventions over time. For pull requests, internal workflows deploy `/ultra review`, spinning up parallel agent reviewers to verify findings and filter false positives before manual review. For ongoing tasks, the system provides both local `/loop` scheduling via cron and server-side 'Routines.' Unlike local loops that require an active machine, Routines run remotely on Anthropic infrastructure triggered by cron schedules, webhooks, or API endpoints. To support growing tool libraries without saturating active context windows, a dynamic 'tool search' indexes and retrieves tool definitions on demand. Tsai notes that alongside enterprise updates—such as expanded Windows support, native cloud provider packages, and centralized admin policies—future iterations of background routines are expected to move beyond compiling issue reports to autonomously deploying subagents to fix identified bugs.

Anthropic技术团队成员Dixon Tsai表示，Claude Code近期更新集中于“提升开发者体验”与“增强自主运行能力”两条主线，旨在优化全天候开发交互并使模型能够独立承担长时间运行的复杂编程任务。该产品通过重构终端与图形界面、引入安全权限评估、支持Git工作区隔离以及构建项目级持久记忆，推进了从本地交互到云端后台自动化的协同开发体系。 交互终端改造与多视图桌面界面 (0:00) 在日常交互环境优化方面，Claude Code推出了“远程控制”（Remote Control）功能。开发者在桌面终端触发该命令后可生成网页链接，通过手机浏览器或移动应用实时查看终端会话并输入指令，从而便于在外出时继续调度子代理处理多PR长会话任务。针对长期存在的文本追加渲染问题，新版采用全屏虚拟化滚动模式替代传统追加方式，在消除终端画面撕裂与重绘闪烁的同时稳定了内存占用，并新增了可折叠工具调用、平滑回滚与空格键语音输入（`/voice`）等交互细节。 桌面图形界面（GUI）同步完成了架构更新以支撑多会话管理。侧边栏新增了按项目分组、筛选以及拖拽分屏并排显示的能力。在对话记录之外，桌面端集成了支持行内步骤标注的人工规划视图（Plan）、支持逐行审查的差异视图（Diff）与全量目录树视图（Files），并引入了可为长对话自动提炼语义化标题的实验性目录索引（Pin as chapter）功能，协助开发者结构化跟进长周期任务进度。 自动化权限评估与并行分支隔离 (10:00) 针对模型在无人值守执行时频频因权限确认弹窗或分支冲突而中断的难题，Tsai展示了“自动模式”（Auto Mode）与Git工作区（Git Worktrees）的协同方案。自动模式内置安全分类器，负责在后台实时研判工具调用是否具备破坏性或存在提示词注入（prompt injection）特征；通过安全检测的操作可免除人工逐次授权直接放行，减少长任务对开发者即时响应的依赖。 为避免并行运行的智能体相互污染代码库，Claude Code接入了Git工作区环境隔离。系统允许开发者在配置中指定共享 `node_modules` 等重型依赖目录，克服了传统工作区重复克隆大体积文件的弊端。在此机制下，Claude可自主调用内置的进入（`enter work tree`）与退出（`exit work tree`）指令，独立在隔离分支中完成功能开发、测试并在交付后清理环境，支持多任务无冲突并发推进。 项目持久记忆与多智能体代码评审 (15:00) 针对每次新建会话时上下文被清空导致的Token浪费与重复引导问题，Claude Code增加了“自动记忆”（Auto Memory）机制。该功能依托Claude 3.7 Sonnet等模型的长文本与目录推理能力，在项目本地自动维护 `memory.md` 索引文件，长期记录团队的构建脚本、调试经验与代码规范，并通过渐进式披露机制按需关联深度文档，确保跨会话认知得以保留。 在代码质量保证层面，Anthropic公开了内部采用的多阶段多智能体代码审查模式（Multi-Agent Code Review）。通过运行 `/ultra review` 或配置GitHub应用，系统能够并行调度多个聚焦不同审查维度的代理角色，交叉验证各自的审查结论并过滤误报噪音，在人工介入复核前提供更高密度的PR诊断结果。 定时循环调度与云端后台例行任务 (20:00) 在任务的主动触发与后台运行方面，Claude Code新增了本地会话内的定时循环（`/loop`）功能，支持模型通过内置的 `cron create` 工具自主按周期执行重复性任务，并能通过自然语言指令停止调度。为防止工具定义过度占据活跃上下文，系统新增了工具检索（Tool Search）机制，允许模型按需动态检索并加载工具配置，从而压减初始上下文负荷。 此外，处于研究预览阶段的“远程例行任务”（Remote Routines）将执行环境转移至Anthropic云端服务器，支持依据Cron定时器、GitHub Webhooks或API端点无须保持本地设备开机持续运行。Tsai展示了该功能每日自动检索特定标签的开源issue并生成权重摘要的实际用例。结合企业级需求，团队还补充了Windows系统支持、云提供商原生部署以及集中式管理员策略配置。 Tsai的演示表明Claude Code正逐步由单一的即时代码生成工具转变为具备自主上下文管理与隔离执行能力的开发平台；不过，诸如远程例行任务等高级自动化功能目前仍处于研究预览阶段，且由自动化摘要向端到端派发子代理自动修复漏洞的闭环链路仍属前瞻设想，实际落地效果仍有待后续迭代验证。

[0:00] Dixon Tsai: remote control [0:00] Dixon Tsai: 3 plus 3 [5:00] Dixon Tsai: 2e full screen [5:00] jump to bottom [5:00] pin as chapter [10:00] Dixon Tsai: Auto Mode [10:00] Dixon Tsai: clod [10:00] enter work tree [10:00] exit work tree [15:00] Dixon Tsai: Auto Memory [15:00] Dixon Tsai: /ultra review [15:00] Dixon Tsai: Routines [20:00] Claude: why do programmers prefer dark mode? Because light attracts bugs. [20:00] Dixon Tsai: stop the loop

[0:00] please welcome to the stage member of technical staff philanthropic Dixon Tsai. All right. Good morning, everyone. I'm Dixon from the Cloud Code team. I work on plugins and Cloud Code on the web. This session is about what's new in Cloud Code. If you've seen our public changelog recently, you'll know that our team ships a lot. It could be kind of hard to keep up. So today, I want to step back from the changelog and give you the story. How our new features fit together and why we love them. Our new features fall into two themes. There we go. Our new features fall into two themes First developer experience How we making Cloud Code nicer to live in as you spend more of your day with it Second autonomy. How we're enabling Cloud to do more while you're away from your keyboard. Finally, I'll cover how to learn more for everything I don't get to. As daily users of Cloud Code, we're still very invested in the human experience, even as Cloud can handle more long-running tasks on its own. For developer experience, first up is remote control. Remote control lets you start a session on your machine and pick it up on the go. From your phone, you can access the same session, same dev setup. Why not kick off a long-running task and run some errands? We wanted more convenience on our team when building Cloud code, and remote control gives us that. Let me show you how to set up remote control with a demo. Alright, so we are in a new session. Let's just seed it with some content. And now we can enable remote control by typing remote control What you get is a link in Cloud Code on the web so let open that and put these side by side and you can see the content of your terminal online And from you know this can be your mobile browser, this can be your mobile app, you can type 3 plus 3 or something like that, and it'll appear in the terminal. Same dev setup that's running in your terminal, You can also type in your terminal and see it in Cloud Code on the web. So my favorite pattern here is to enable remote control on a long-running session that has accumulated context over multiple PRs. That way, when I'm on the go, I can ask that long-running session to spin up a subagent and get my idea implemented. Back to the slides. Next is flicker-free rendering. Show of hands, how many of you have encountered flickering at some point when using Cloud Code? That's how you know that you're an OG. So we heard from you loud and clear that flickering ruins the terminal experience Our previous approach of appending to the existing terminal scrollback was tricky to get right One misalignment could trigger repainting that would cause the flickering that you see. Now, Cloud Code's terminal UI has fullscreen mode, where we virtualize scrollback. By doing so, we can guarantee flicker-free output. On top of that, we can make elements of quad code in the terminal clickable. And we can keep memory usage flat, even for very long sessions. Our team has switched to this for the most part, and you'll feel the difference immediately if you do as well. Let me show you how to use it in a demo. So, in a new session, I'll expand this. You can enable full screen mode by saying 2e full screen. 2e is short for terminal UI. By the way, did you know that there's a voice mode in Cloud Code? How many of you have used voice mode already? Okay, a fair number of you. You can toggle it on or off by using slash voice. And now you can hold the space bar to say a prompt. [5:00] So I'll say, write me a file containing 10 jokes and another file containing 10 fun facts. Containing 10, oh, there we go. All right, so as cloud is working, you'll see how smooth the rendering is for full screen mode. Even for very long content, you don't see any rendering artifacts. And one thing you can do is for these tool calls that are folded, you can expand them with a single click. Now let me show you how virtualized scrolling works with another prompt. Create a file containing 10 career tips. Career tips. Okay, there we go. And as it's working, let's say you are scrolling back and reviewing the existing output. We see this jump to bottom here that you can click and our existing prompt does its processing and has alerted us that there's a new message. So it doesn affect your view above the transcript So we can then jump down and see what that new message is This is full screen mode So if you like me and you spend a lot of time in the terminal full screen mode provides you a really smooth intuitive experience. Back to the slides. Some engineers on our team prefer the terminal, but others are increasingly using the cloud code GUI across desktop and web. We've revamped the GUI into a much better experience for managing, one, multiple sessions at once, and two, presenting all of the relevant context you need for the session so that you can work more efficiently. Let's take a look at some of these new features in a demo. This demo is based on Excalibur. So Excalindraw is an open source whiteboarding app and the original experience has three basic shapes, rectangles, diamonds and ellipses. But let's say I wanted to draw trapezoids. Well, I have a Cloud Code Desktop session cooking with this feature. If you haven seen Cloud Code Desktop in a while a lot has changed In the sidebar we have a pinned section and also more advanced filtering and grouping capabilities In fact you can group by your project for example From the sidebar, you can also open up a split view, just like that, with drag and drop. On the top right corner, you can see a lot of views for your session, because the transcript's kind of boring sometimes, right? You can open the plan corresponding to the session. You can even highlight certain parts to leave comments, and all the comments can get aggregated and resolved by Claude later. So in addition to the plan, you can open a diff view and also leave comments on any line. Finally, if you're kind of lost, you can open the files view where you have your entire working directory and you can also make edits really quickly. But what we're betting here is that you're spending more and more of your time with Claude as Claude becomes more capable and all these functionalities are more optional, right? They're not your default experience, they just something you grab onto when you need them Now let me show you this really cool experimental feature So you know that this transcript is really long Now one thing we can do is we can hover over an assistant message and there's a pin as chapter button. You should try to guess what this does. So let me click here and then you can see that above this assistant message we now have a title for this message in the context of this transcript. You can rename this title debug pointer down event. Let's maybe add one more. How about this one? So we'll pin this as a chapter 2 and see what its title is. The title is add missing distance to trapezoid element function. Cool. So we now have two titles. What can we build with this? We can build a table of contents. So now in the top left corner you can jump between different parts of your transcript. You can rename these chapter titles or you can even delete them if you want. So that's an experimental feature and we think that that we can make a lot more on Cloud Code Desktop. [10:00] It's Anthony Morris, our Cloud Code Desktop lead, No on X, and he'll probably ship it for you tomorrow. So one last thing here is we haven't looked at trapezoids, so let's actually run a server. We can open up the terminal right here, start the server. It's going to run in localhost 3002, and now you can draw trapezoids, and you can even make them orange. All right, let's go back to the slides. So that's for developer experience. Now on to our second theme, autonomy. Each of these next features came from the same place. I'd hand Claude something long running and one small thing, a permission prompt, a branch conflict, a forgotten build command would lead me to wonder, Claude, why'd you do that when I come back to my laptop? The features I'll cover next will give you more confidence that Claude will do what you want when you step away. First feature I'll cover, show of hands. How many of you have become tired of permission prompts? Where you go one one one You can probably start a wave right With your hands Auto mode is a new permissions mode in Cloud Code where Cloud makes permission decisions on your behalf using a classifier So when you have a tool call that would otherwise hit a permission prompt, the classifier will look for two things. One, is the action destructive? And two, does it look like prompt injection? If the classifier determines a tool call it's safe, it just lets it run. If not, then Claude will block the tool call and the agent will have to figure out what to do next. This has saved us a bunch of time at Anthropic and helps us eliminate the number of instances where we thought we started a long-running task just to come back and be like, ugh, a permission prompt. So next is, oops, sorry. Next is Worktrees. Next pain point is how do we work with multiple features without your clods stepping on each other? Work trees are my favorite feature because they've helped me overcome this pain. Git work trees have existed long before clod code where it will create a new branch and it will copy all the files of your repo But there are so many sharp edges and so we taken the time to think about what those edges are and to smooth them out into a friendlier interface for us and for Cloud. So, for instance, we've integrated in Cloud Code settings a field where you can declare the files that should be shared across your work trees, like Node modules. So then, if you wanted to start a new session, you can just use dash dash work tree, and that new session will be accompanied by a new work tree. Or you can ask Claude to create a work tree for you. Let's take a look at this in a demo. So again, our demo is based on Excalibur. In the original experience, you see here, we're going to add two things for our demo. Now, one is we're going to add a new color, and two, we're going to add a slider to help us make these corners more round. So we going to start a work tree by using clod which is short for dash dash work tree So we have a work tree for colors a work tree for border radius and I just gonna write in the prompt right now add a sixth color to stroke slash background And then for this work tree, I'm gonna say add a slider to adjust border radius. Alright, so as these two sessions are cooking their respective features, Let me show you how Claude can create work trees on its own. So, I can say for instance, create a work tree to add a star shape. And so, it's not just dash dash work tree, we've actually given Claude an enter work tree and exit work tree tool so Claude can determine for itself whether it should start implementing a new feature in an isolated environment. So you can see that Claude has created a work tree. I'm being explicit in this demo, but you can put this prompt in your ClaudeMD. You don't have to be as explicit. Claude will know what your intent is. And Claude is polite enough to ask me when to start implementing. I'll actually say just leave the work tree. [15:00] see that our harness cleans up the work tree for you. So it's very convenient. And that way you can have a long running session that works on many different PRs simultaneously. So let's take a look at our features now. Looks like the colors are ready. So let's run yarn start in the background and tell me the port. The other feature is still working. So this will take a little moment. All right, so this is ready in localhost 3003. And you can see I still have the original experience running on my main checkout of Excalibur. This new experience has the sixth color. So this feature, the agent was working independently on this feature. And then we can see how this other one is doing. It's still working, but I think you get the idea. Work trees are just super convenient for launching multiple agents to work on multiple features Back to the slides Next is auto memory How many of you feel like you spend a lot of tokens on the same things over and over again because you think Cloud Code's working on a blank slate, right? But with auto memory, we now let Cloud accumulate knowledge across your session so you don't have to set it up yourself. You can ask it to remember things like key build commands, debugging insights, and any other project preferences. And Claude will decide what's worth remembering based on whether the information could be useful for a future conversation. In particular, Claude for every project will manage a directory and within the directory is a memory.md file. It's an index file that gets included in your conversation just like claude.md but it's exclusively managed by claude and from this index file it can offer more files to... sorry, it can link to other files in the same directory that have more details So we leverage progressive disclosure much like skills so this feature is very handy it also works for sub so if you have a custom sub agent you can say memory user or memory project to establish a memory directory and this feature wouldn't be possible without recent breakthroughs in model development so compared to opus 4 opus 4.7 and other recent models are more capable of managing its memory as a set of files in a directory. So if you've been following along, Claude's been doing much better at Pokemon because of this capability. Next up is code review. At Anthropic we have discovered a more effective code review pattern that's multi-phase and multi-agent. Spin up a team of reviewers to look at independent aspects of your code, then verify that that the findings are actually worth reporting. This process catches issues that would have otherwise taken you hours or that you would have missed altogether. So you can set up this code review for all of your PRs through our GitHub app or you can try it out manually in cloud code by using slash ultra review So between auto mode work trees auto memory and code review you now have fewer actions to approve more agents safely working in parallel sessions accumulating knowledge over time and teams of Claude reviewing changes all before you yourself have to look at each PR your time is valuable and we should leverage Claude's ability to make the most of your time all right and that brings us to the final feature we're covering today routines routines have shipped in research preview recently and they help you create sessions in Claude code that can trigger without your intervention so we've talked about how we have features that help the agents work more autonomous from idea to PR and now we're automating the idea generation itself, right? The shape of a routine is simple. You configure them once with your prompt, your repo, and relevant connectors, and then you can select a trigger, either a cron schedule, something that runs hourly or daily, a GitHub webhook event, or you can even set up an API endpoint that you can use. [20:00] for the routine to run. From there, Cloud runs without you. So you can have a routine that triages GitHub issues every day. And then if you enable the Slack connector, you can have it send a digest of those top issues to Slack. Or you can have a routine that gets run every time your e-commerce website earns a sale. You can have Cloud be happy with you as well. So let's see how routines work with a demo. Now, routines are a feature that is very prominent now in the sidebar, but we also introduced slash loop. So let's start a new session. What we can say is slash loop one minute, tell me a joke, for instance. And then I can use my backup side by side. So, you know, split view is kind of nice. So, slash loop can help you set up automation within a single Cloud Code session. So Cloud has a cron create tool similar to the enter work tree tool where it can at its discretion start a scheduled job for you and of course you can always disable this tool in your settings if you don like it so now I've had it scheduled tell me a joke every minute and so there's a joke about why programmers prefer prefer dark mode because light attracts bugs etc etc. So compare this against routines where you can kick off a new routine either locally on your desktop or remotely. I prefer remote because if you set up a routine locally you have to have your laptop open but if you run it remotely our servers will schedule and run these jobs for you. In fact I have my own routine I've been running a lot recently. Let's walk through it. So I have my instructions here are scan the latest open public GitHub issues with the most reacts created since the day before. Look only at the issues with labels and area plugins and area hooks, which I'm responsible for, and respond with a separate list of the top 10 issues sorted by the most reacts. And so it been running for a week now and you can see that every day I get a really nice report of the top issues and I don have to think about prompting Claude to launch these reports And you can imagine we can set up other automation to then ingest these reports, launch subagents to fix them, etc. Let's go back to our jokes. So once we're done, you know, writing jokes, we can say stop the loop. And Claude can interpret your prompt, map it to the proper tool call, and delete it. So you'll also notice that we have a tool search tool now. What tool search does is it adds a layer of indirection between you and the actual tool definition. So now we can offer more tools at once without directly impacting your context. All right. Back to the slides. So we've covered developer experience. We covered autonomy. Now let go through all the other items that I couldn get to in this conversation You can take a picture or you can just get a sense of the overwhelming number of features our team ships every day in particular for those of you Managing cloud code for your team or enterprise we've built some features for you For example we have much better windows support than we've had months before we have cloud provider setup native installation, admin settings management, and plugin features for company-wide harness improvements. As always, you can follow along in our public change log. Beyond the change log, we have a few ways to keep up with what's new. You can follow Claude Devs on X, you can read our weekly What's New Digests in our documentation, or you can subscribe to our developer newsletter. We hope you get a a chance to try at least one of the features that we covered today and let us know what you think. And then for those of you in the room with us, we'll be having a conversation with Dario and Daniela, moderated by Ami, right here on the main stage at 1 p.m. Thank you very much, everyone.