Claude Tag：主动记忆并自主行动的 AI

如果频道里没有 Claude 怎么办？

What do you do if there's no Claude in the channel?

我不知道。我记不得没有 Claude 在频道里的世界了。

I don't know. I can't remember a world without Claude in the channel.

带我回顾一下。三年前，我们刚开始把 AI 融入日常工作流程。那时候是什么样子，一直到现在又经历了什么？

Walk me back. So, 3 years ago, we were just starting to adopt AI into our day-to-day workflows. What was it like back then and take me all the way to where we are today?

我记得两年前，我还在用代码补全。那大概就是 AI 作为工程师能做的事。后来我们构建了 Claude Code，引入了智能体，以及新一代的编程方式。从那以后变化太大了。比如，用代码补全时，我还是在决定要写哪行代码，智能体只是稍微帮点忙，只是补全那一行。然后我们发展到它能写整个函数、整个文件、整个功能。我觉得现在有了 Tag，它直接搞定一切。它能写功能，能做完整的端到端实验，能做我的数据分析。这就像两年内两次飞跃。回想过去几年，几乎就像一次转型。我们从一个人始终在循环中，一个人坐在那里一次只打一行代码，变成一个人坐在那里带着大概 10 个 Claude，一次打一堆功能。现在我们到了 Claude 实际上更主动驱动的阶段。而且不再只是一个人了，而是整个团队在与它互动。

I remember 2 years ago, I was still using type ahead. That was like what AI could do as an engineer. And, you know, we built Claude Code, we brought agents and kind of the next generation of ways to do your coding. And it's just changed so much since. Like, you know, with type ahead, I was still deciding I want to write this line of code and the agent was helping a little bit. It was just spreading the line. And then we got to the point where it was writing whole functions, whole files, whole features. And I feel like now with Tag, it just does the whole thing. Like, it can write the feature, it can do an entire experiment end to end, it does my data analysis. It's been like two leaps in 2 years. When I think back the last couple of years, it's almost like this transition. We went from a person always in the loop, it's one person sitting there and they're just typing a line at a time. And we got to one person sitting there with like 10 Claudes, you know, typing a bunch of features at a time. Now we're at the point where Claude is actually driving a little bit more. And it's not just one person anymore. It's, you know, an entire team that's interacting with it.

那么，什么是 Claude Tag？

So, what is Claude Tag?

过去，你必须打开 Claude，问它点什么，然后 Claude 才会干活。它会用你的工具，用你的电脑，完成任务。以前就是这样。而有了 Claude Tag，Claude 会主动介入。它是主动的，知道什么时候该介入。即使任务需要几天或几周，它也会完成并跟进。我觉得最酷的是它会记住我告诉它的内容，以便下次使用。以前我每次都得提示 Claude，说‘好了 Claude，现在做这个，现在做那个’，我还得记得去提示。现在我们只需把 Claude 加入频道，它就会主动介入并完成工作，每个人都能看到，每个人都能参与，因为它是多人的，而且它会记住下次怎么做。

In the past, you had to open Claude and you had to ask it something and then Claude will do the work. And it'll use your tools, it'll use your computer, it'll get the task done. That's the way that it used to be. With Claude Tag, Claude jumps in. It's proactive. It knows when to jump in. It'll do the work even if it takes days or weeks and it'll follow up. And I think the coolest thing is it'll remember what I told it for next time. So before I would have to prompt Claude every time and I would have to say, 'Okay, Claude, now do this or now do this.' I would have to remember to do that. Now we just add Claude to a channel, it will proactively jump in and do the work, everyone gets to see it, everyone gets to participate because it's multiplayer, and it'll remember for next time.

好的，那么研究方面是什么真正促成了这一点？Tag 为什么这么好？

Okay, so what on the research side actually enabled this? How is it that Tag is so good?

多年来我们一直在努力让模型更持久、更自主。如果你看看最新的评估，我们最新的模型一次能工作 16 小时，现在已经到了我们甚至无法准确检测它能工作多久的程度。在 Claude Tag 这样的东西里，Claude 能够自我安排工作。所以你可以把这个 16 小时的任务，实际上它还会随时间增加，因为 Claude 可以在几天、几周或几个月后安排后续任务。我觉得以前我们不敢让它做的任务，现在它几乎每次都能完成。

For many years we've been working on making our models more long-running and autonomous. So if you look at the latest meter evals, our latest models can work for 16 hours at a time and are now in the zone where we can't even accurately detect how long it's able to work for. In something like Claude Tag, Claude is able to self-schedule work for itself. And so you can take this one 16-hour task and it actually increases over time because Claude can schedule a follow-up after days or weeks or months. I feel like the task that it we couldn't trust it to do before now it can get almost every single time.

我个人就有很多 Claude Tag 会话，运行几天、几周。我想我可能有一个运行了一个月的，基本上就是一个长期运行的实验，它每天检查数据，偶尔发送一个 bug 修复，如果有 bug 需要修复的话。我就看到拉取请求进来，每天收到这些数据报告。

I have a bunch of these Claude Tag sessions personally that are running for like days, weeks. I think I might have one that's like a month and it's essentially this long-running experiment and it's just checking in every day, checks the data, once in a while it sends a bug fix if there's some bug that needs to get fixed. I just see the pull request coming in and I get these data readouts every day.

另一个重大创新就是模型能够拥有记忆。所以模型不仅能给自己设置提醒，还能记住所有用户随时间给它的所有指令。

The other big innovation is just being able to have memory in the model. So the model can not only set reminders for itself, but it also remember just all the instructions that all the users have given it over time.

记忆功能花了很长时间才攻克。我觉得我们为 Claude Code 尝试了好几年，现在感觉终于做对了。用起来感觉非常好。

Memory took a long time to crack. Like I think we tried to get away for Claude Code for years and it feels like we finally got it right. It feels really good to use.

好的。所以，如果你在频道里使用 Claude Tag，告诉它‘嘿，我只希望你监控这类问题，而不是其他类别’，它会照做。它会永远记住那个频道的要求。然后如果其他人说‘嘿，实际上，让我把你的范围扩大到这件新事’，它会调整。另一个大家喜欢的特点是 Claude 的个性和情商。有时我和人们讨论把 Claude Tag 加入频道时，他们会说‘哦，我怎么知道它不会在每个线程里都插嘴，变得烦人？’实际上，Claude 经过训练，对何时需要它有很好的判断，它也可以退居幕后。如果它走偏了，你可以直接告诉它做不同的事，少介入或多介入，它会记住以后怎么做。

Okay. So, if you're working with Claude Tag in a channel, and you tell it, 'Hey, I want you to monitor only for this type of issue, but not these other categories.' It'll do that. It'll remember that for that channel forever after. And then if someone else says, 'Hey, actually, let me expand your scope to this new thing.' It'll adjust. Another big thing that people love is Claude's personality and EQ. Sometimes when I talk with people about adding Claude Tag into a channel, they say, 'Oh, how do I know it won't jump in on every thread and be annoying?' Well, Claude is trained to have a good sense for when it's needed, and it can take the backseat. And if it ever goes too much in the wrong direction, you can just tell it to do something different, to jump in less or jump in more, and it'll just remember it for the future.

那么，当你观察内部使用情况，以及和刚开始使用 Claude Tag 的客户交流时，你看到了哪些模式？

So, when you look at the internal usage, and when you talk with our customers who are just getting started on Claude Tag, what are the patterns that you see?

对我来说最酷的是客户如何摸索出用法。它可以写你的拉取请求，可以调试生产问题，可以做数据分析。你只需要把它连接到你的工具上。我认为这之所以可能，完全是因为记忆功能太好了。我看到人们做这种涌现出来的事情，比如，这是一个问答频道。每当一个问题被回答，你必须打勾确认。你告诉 Claude Tag 这个规则，它就会照做。它会回答问题，然后每次用勾选表情回应，就这样为你做。比如我们有一个数据问题频道，现在 Claude Tag 就回答所有问题。或者我们有一个 Claude Code 反馈频道，Tag 现在修复一切。

The coolest thing to me has been how customers are figuring it out. It can write your pull requests. You can debug production issues. It can do data analysis. All you have to do is hook it up to your tools. I think that's only possible because the memory is just so good. And I've seen people do this kind of emergent stuff where, you know, for example, this is a channel for Q&A. Anytime a question is answered, you must check it off. You tell Claude Tag this, and it'll just do that. It'll answer the questions, and then it'll react with a check every time, and then it'll just do this for you. Like we have a channel for data questions, and now Claude Tag just answers all the questions. Or we have a channel for Claude Code feedback, and Tag just fixes everything now.

我问数据问题最常见的方式就是发在我们的数据频道里，然后 @Claude Tag，它先回答第一轮。通常效果不错。

The most common way that I ask a data question is I post in our data channel, but then I tag Claude Tag, and then it takes the first pass. And normally it's pretty good.

我在一堆频道里一直这样做，后来我厌倦了 @它，就直接告诉它‘请每次都回复’。

I kept doing this in a bunch of channels, and then I just got tired of tagging it, so I just told it, 'Please respond every time.'

我认为这正好展示了记忆的强大，你能做到这样的事。你只需告诉它‘嘿，记住总是做这件事’，它就会支持你。这极大地改变了我们所有人的工作方式。我认为这也是一种让更多人理解如何与 AI 协作的方式。因为 Claude Tag 在公共频道中工作，每个人都能观察到专家用户如何利用它。然后他们把这些模式带到自己正在做的新项目中。所以无论是在内部还是在我们合作的客户中，我们都看到了 Claude Tag 最佳实践的扩散，我认为这对 AI 工具来说非常新颖。

I think it was to show how powerful the memory is that you're able to do something like this. You can just tell it, 'Hey, remember to always do this thing.' and it just has your back. It's been a huge shift in how we all do work. I think this also is a way for more people to understand how to work with AI. Because Claude Tag is working in public channels, everyone is able to observe how the expert users leverage it. And then they take those patterns and they bring it to new projects that they're working in. And so both internally and in the customers that we've been working with, we've been seeing this diffusion of Claude Tag best practices that is I think very novel to AI tools.

它接管的速度之快令人惊讶。我觉得现在我在我们所有的频道里都能看到 Tag，比如每个反馈频道、每个数据频道。一开始只有几个人在用，但我觉得当其他人看到那些人使用时，他们很快就学会了。

It was surprising how fast it took over. Like I feel like now I see Tag in just every channel that we have, like every feedback channel, every data channel. It started with just a few people using it, but I think as other people saw those people using it, they really quickly picked it up.

拥有 Claude Tag 的支持真的让人很有底气。我们现有的很多工具，比如聊天、代码工作和 Claude Code，都需要你记得主动打开它们，所以体验上更偏向被动响应。Claude Tag 很棒的一点是，你给它一个更高层次的目标，比如“为这个频道里的每个 bug 提交 PR”，它就会默默支持你，为每一个 bug 提交 PR，完全不需要你提醒。第二个很酷的地方是它支持多人协作。现在人们使用的几乎所有 AI 工具都只是你一个人和 Claude 合作，然后把输出复制粘贴给你的团队。我认为 Claude Tag 带来的重大转变是，它把 Claude 直接带到了你的工作中心，让多个人可以共同引导会话，得到更好的输出。这里面有很多动态变化：比如，我们通过观察别人来学习如何更好地使用 Claude Tag；Claude Tag 通过听取我们给它的指导来学习如何更好地与我们合作。最终，客户能得到更好的结果，因为这不只是我一个人的意见，而是我们整个团队都能参与进来，把 PR 调整到最佳状态。

It's been very empowering to be able to have Claude Tag have my back. So for a lot of our existing tools, like chat, code work and Claude code, you typically have to remember to open it. And so it's a more reactive experience. And one thing that's great about Claude Tag is you give it this like higher level objective, like put up PR's for every bug in this channel, and it just has your back and puts up a PR for every single one without you needing to remind it. The second thing that's really cool is that it's multiplayer. So almost all of the AI tools that people use right now are just you and Claude working together and then copying and pasting this output to your team. And I think the big shift that Claude Tag allows is it brings Claude right into the middle of your work so that multiple people can guide the session to a better output. Like there's so many like dynamics here. Like one, we're all learning about how to use Claude Tag better by observing other people. Claude Tag is learning how to better work with us by hearing the guidance that we're giving it. And at the end of the day, the customer gets a better result because it's not just my opinion of how we should solve it. It's our whole team is able to jump in to like nudge this PR to the best possible state.

我很喜欢这个想法。我一直告诉大家，我希望每个人都能为代码库做贡献，但很多人害怕，因为他们需要打开终端，甚至是桌面应用，还要处理 Git 和代码检出之类的事情。有了 Tag，他们就能直接做了，真的很酷。

I mean, I love this idea. I keep telling everyone I want everyone to contribute to the code base, and so many people are afraid because they have to like open a terminal, even the desktop app because you have to like deal with a Git and like code checkouts and stuff. And with Tag, they can just do it. It's really cool.

另一件我认为非常重要的事情是，确保你的协作平台有公共频道。这样 Claude Tag 就能监控你正在跟踪的不同项目。作为一名产品经理，我经常同时处理五到十个功能，我会让 Claude Tag 查看每个功能的状态，然后给我一份每日报告。我们看到的另一个重大突破是，人们现在可以自助解答问题。当新人加入我们公司时，他们不用去问法务能不能说这个，或者问人力资源团队我们的福利是什么，他们只需要 @Claude Tag，因为 Claude Tag 连接了我们的真实数据源，无论什么时间，它都能给出非常快速的回答。

The other thing that I think is really important is making sure that your collaboration platform has public channels. That way, Claude Tag can monitor different projects that you're tracking. So, as a PM, I'm often working on maybe five to 10 features, and I have Claude Tag look at the status for every single one and then give me a daily report. The other big unlock we've been seeing is that people can now self-serve their questions. When people onboard to our company, instead of asking legal about whether they can say this or instead of asking our HR team what our benefits are, they can just tag Claude Tag, and because Claude Tag is connected with our source of truth files, it can give them a really fast answer no matter what time they're looking for it.

最大的变化就是它让每个人的效率都大大提高了。这得益于我们内部的 Claude Tag 版本。只看产品部门，也就是我们所在的这部分，由 Tag 编写的 PR 数量现在已经达到 65% 左右，而且还在不断攀升。我之前以为 Claude Code 是让工程师更快的东西，但这个产品让工程师快得多。我认为很大一部分原因在于交互方式：你让 Tag 做一件事，它就会响应，然后你就可以继续下一个任务了。所以，你实际上希望设置成 Tag 能自己完成工作，而你不需要一直检查进度。

The single biggest thing is just how much more productive it's made everyone. And, you know, this is driven by our internal version of Claude Tag. And if you look at just the product org, so, you know, the part of the company that we sit in, the number of PRs that are written by Tag, I think it's like 65% now, and it's just climbing like this. Like I thought that Claude Code was the thing that makes engineers go faster. This is a product that makes engineers go way faster. And I think actually a lot of the reason is because the way you're interacting with it, you ask Tag to do something, and it'll respond, and you've moved on to the next task. So, you actually want to set it up in a way where Tag can do the work and you don't have to check in all the time.

那么，你运行着数百个 Claude Tag，但仍然是 Claude Code 的重度用户。你如何决定什么时候用哪个呢？

So, you have hundreds of these Claude Tags running, but you're still a power user of Claude Code. How do you decide when to reach for each of them?

是的，我以前所有事情都用 Claude Code，后来开始越来越多地用 Tag。一开始只是非常简单的修复，比如有人报告 bug，按钮偏移了几个像素之类的，我就让 Claude 去修复。或者有人问一个很简单的问题，我也让它去处理。过去几周，随着我对产品越来越熟悉，我开始用 Tag 处理越来越多的事情，甚至更复杂的工作它也能完成，因为它可以验证自己的工作——它运行在和我们移动端、桌面应用相同的远程沙盒中，使用相同的智能体 SDK，所以同样智能。当你给它添加记忆后，在不同的频道里，我对验证方式有特定的偏好，每个频道可能略有不同。通常它会修复 bug，然后在 Slack 里发回一个视频，我甚至不需要离开。好了，Claude Tag 下一步有什么计划？

Yeah, so I used to use Claude Code for everything and then I started using Tag for more and more things. First, it was really simple fixes. So, you know, someone has a bug, the button's off by a few pixels or something. I'll just ask Claude, 'Please fix it.' Or someone has a really simple data question. I'll ask them to fix it. And I think what's been happening more and more over the last few weeks is I've been getting more comfortable with the product is I've been using Tag for just more and more things. And even more complicated work it's actually able to do because it can verify its work because it's running in, you know, the same remote sandbox that we use for mobile and for the desktop app and it's using the same agent SDK. So, it's just as intelligent. When you add memory to this, in different channels I have actually specific preferences about how I want it to verify. It might be a little bit different for every channel. And so, often what it does is it'll, you know, fix some bug, then it'll post back a video in Slack and I don't even have to leave it. Okay, so what's next for Claude Tag?

我们在 Slack 上推出了 Claude Tag，并且很高兴能把它带到更多人们协作的平台，比如 Microsoft Teams。我们想确保每个知识工作者无论在哪里工作，都能随时用到 Claude。我认为 Claude Tag 最酷的一点是，它不仅改变了个人工作方式，还把 Claude 置于团队中心，从而改变整个组织。我们让 Claude Tag 像 Claude Code 一样高度可定制，我很期待看到各个组织如何定制 Claude Tag，让它成为自己的工具。

So, we launched Claude Tag in Slack and we're excited to bring it to more platforms where people collaborate like Microsoft Teams. We want to make sure that every knowledge worker is within arm's length of Claude no matter where they're getting their work done. And I think one of the coolest parts of Claude Tag is it's not just changing how individuals work. It's putting Claude in the center of teams and transforming entire orgs. We built Claude Tag to be incredibly customizable just like Claude Code and I'm so excited to see how orgs customize Claude Tag and make it their own.

我已经等不及了。

I can't wait.