Anthropic 技术团队成员:当 Claude 从助手变成主动协作的 AI 队友
原题Member of Technical Staff at Anthropic | When AI Joins the Team
推荐理由Anthropic FDE 团队负责人用项目追踪、反馈分流与排班协调,讲清 Claude 如何逐步接过持续职责。希望从个人提效走向团队协作的读者,可以学习共享上下文、工具身份、记忆、触发和反馈怎样配合,并保留人对需求、代码与行动项的判断。
看前速览
Anthropic FDE 团队负责人介绍,如何借助逐步授权、共享上下文、工具、记忆、触发机制和反馈,让 Claude 成为主动协作的队友,同时保留人的关键决策。
通过小步交接逐渐扩大职责
团队从问答开始,逐步让 Claude 创建、维护跟踪工具,再运营接收用户反馈的内部应用。每一步都针对已经出现的瓶颈,而不是先规划宏大的全自主系统。
区分自主维护与产品决策
Claude 分类反馈、修复明显缺陷,人则决定要开发哪些功能并评审 PR。把职责交给系统,并不意味着所有判断和审批环节都会消失。
让 AI 队友主动向人索取缺失信息
每周五收集项目更新、核对上下文,为周一排班会议做准备。无法解决的问题进入议程,而不是让所有负责人一起重新拼凑当前情况。
用人工确认闭合行动循环
会议转录先变成待审核行动项,人确认后,Claude 再协助更新跟踪表、发布消息和通知团队,让协调真正转化为执行。
五项基础能力支撑主动协作
共享上下文、带审计记录的工具权限、分层记忆、主动触发和能够修改指令的反馈,让 Claude 在团队已有工作界面中协作并积累有用知识。
完整摘要
Anthropic 一位前沿部署工程团队负责人介绍,他入职后发现同事把 Claude 当成团队共享工作方式的一部分,而不只是个人聊天助手。入职手续、差旅、客户背景等问题,可以通过应用、Slack 或 Claude Code,利用公司知识直接回答。随后,演讲沿着 FDE 团队的成长展示几次职责交接:先让 Claude 创建跟踪表,再让它持续维护新项目数据,最后通过经过人工评审的 PR,把表格替换成内部应用。
用户反馈到来后,Claude 负责分类和修复明显缺陷,人仍决定哪些新功能值得设计和开发。工具完善后,瓶颈转移到协调:一次排班讨论发现,同一工程师分别被分配了 100%、50% 和 25% 的工作量。团队于是要求 Claude 每周五追踪更新,综合频道和记录中的信息,把无法解决的问题整理进周一议程。会议录音转录后,它提取行动项,交由人确认,再协助执行已批准的后续工作。
支撑这些变化的条件包括共享上下文、能够以自身或用户身份调用并留下审计记录的工具权限、个人与团队及组织记忆、时间或事件触发,以及持续更新指令的反馈循环。关键变化是,人不再反复分派孤立任务,而是由 AI 队友主动向人索取缺失信息。案例保留了人工决策环节,但没有提供生产率研究、失败率或详细权限设计,因此应视为第一方实施经验,而非无限制自主运行普遍可靠的证明。
章节
文字稿
简要描述Brief Description
Anthropic 一位前沿部署工程负责人追溯了 Claude 从回答问题,到维护工具并协调团队协作的发展历程,随后介绍了五项关键基础组件。An Anthropic forward-deployed engineering leader traces Claude's evolution from answering questions to maintaining tools and coordinating team work, then explains five enabling building blocks.
目录Table of Contents
- Joining Anthropic and asking Claude
- 从任务到职责From tasks to responsibilities
- 一套适用于不断壮大的团队的系统A system for the growing team
- 人员配置与主动协调Staffing and proactive coordination
- AI 队友的五个构成要素Five building blocks for an AI teammate
Joining Anthropic and asking Claude
我叫 Amandeep,在 Anthropic 负责带领一支前沿部署工程师团队。对我来说,Anthropic 和这份工作都还比较新。我入职大约三个月了。不过在 AI 领域,这段时间感觉要长得多。My name is Amandeep. I lead a team of forward deployed engineers at Anthropic. Both Anthropic and this role are relatively new for me. I'm about 3 months into the job. Which in the world of AI feels way longer.
就在六个月前,我还在 AWS,负责从零到一推进多个项目,打造新产品,其中一些产品你们都已经很熟悉了。自从 ChatGPT 发布以来,大约三年半的时间里,我一直在折腾各种 AI 相关的东西。我搭建过自己的智能体,也用过 Claude、Codex 等常见工具。Now 6 months ago I was at AWS where I led several initiatives zero to one building new products and some products that you already you are all aware of. And I've been tinkering with all things AI for three and a half years or so since ChatGPT came out. I've built my own agents, I worked with Claude, Codex, all the usual suspects.
我原以为自己算是比较资深的 AI 用户。后来我发现并非如此。当我和——当我开始与 Anthropic 的人交流时,无论是那里工作的朋友,还是在面试过程中遇到的人,我反复听到的一句话是:他们说,在 Anthropic,我们使用 Claude 的方式不一样。And I thought that I was a relatively advanced user of AI. And then I discovered that that's not true. When I talked to when I started talking to people at Anthropic both friends who work there as well as through the interview process one thing that I heard consistently was they said that at Anthropic we use Claude differently.
你知道那听起来像什么吗?就像有人说,我们这家 CrossFit 健身房练 CrossFit 的方式不一样。我不相信。但我确实好奇起来,也开始琢磨。我知道他们用 Claude 比我多。You know what that sounded like to me? Somebody saying that we do CrossFit differently at my CrossFit gym. I don't believe it. But I did get curious. I did start wondering. I knew that they used Claude more than I did.
这一点是毋庸置疑的。我不确定的是,他们使用 Claude 时是否有不同的心智模型?他们的系统是否不同?习惯是否不同?加入之后我发现,各个团队组织自身和工作的方式确实不一样。That was a given. What I wasn't sure about was did they have a different mental model for using Claude? Did they have different systems? Did they have different habits? And once I joined I found that the way teams organize themselves and their work is different.
随着 Anthropic 的 FDE 团队不断发展,我得以近距离见证这一切。今天我要讲的,就是 Claude 作为队友的故事。我是六月底加入的,第一周全都在参加入职培训。And I got to witness this up close as the FDE team at Anthropic has evolved. And that's the story I'm going to tell you today about Claude as a teammate. Now I joined in June, end of June. The first week was all orientation.
过了几天,我们拿到了笔记本电脑。我开始进行设置。我们有详细的说明。And a couple of days in, we get our laptops. And I start to set things up. We've got detailed instructions.
一切都挺好。按照那些说明操作,一切都正常。不过,还是会有一些问题冒出来,对吧?我就有个问题:我想安装 Microsoft Word,该怎么做?于是我问 IT 人员:“我该去哪里获取许可证?”他们说:“直接问 Claude 就行。”于是我就这么做了。All good. Following those instructions, everything's working. But again, there's questions that come up, right? So, I have a question. I want to install Microsoft Word. What do I do? So, I asked the IT person, "Where do I get licenses from?" And they said, "Just ask Claude." And so, that's what I did.
我打开应用,直接问:“我在哪里可以下载 Word?”它给了我一个很直接的回答。接着又冒出了更多问题,比如:“我在哪里注册福利?”然后,我就去问了 Claude。I opened the app, went and said, "Where do I get Word from?" And it gave me a response, straightforward. And there's more questions that come up. Where do I sign up for benefits? And then, I went and asked Claude.
后来我又有了一个问题。我有机会去客户现场。这次有人说:“嘿,你可以在 Slack 上给 Claude 发私信,问它一个问题。”于是我就这么做了,对吧?你无论在哪里工作,都可以使用 Claude。Then I had another question. I got to travel to a customer site. This time, somebody said, "Hey, you can DM Claude in Slack and ask it a question." And so, that's what I did. Right? Claude's available wherever you work.
问 Claude 就行。这似乎是我所有问题的答案。Just ask Claude. That seemed to be the answer to every question that I had.
而事实证明,这正是 Anthropic 日常发生的事情。人们总是在各种不同的平台上向 Claude 提出各种各样的问题。Claude 应用、Slack、Claude Code,任你选择。而这些问题的类型也五花八门。And turns out, this is just what happens at Anthropic. People ask Claude all kinds of questions all the time in all different surfaces. The Claude app, Slack, Claude Code, take your pick. And these questions there's a range of these questions.
从“我该去哪里给孩子买连体衣?”到“我该去纽约办公室哪里接待访客?”,以及介于两者之间的一切。我想了解一位我即将合作的特定客户。How do Where do I get onesies for my kid? To where do I get visitors in the New York office? And everything in between. I wanted to learn about a particular customer that I was going to be working with.
而我当时对情况一无所知。那我会怎么做?通常我会去找客户经理说:“跟我讲讲这个客户。讲讲这些项目。进展如何?我们都了解些什么?”And I did not have any context. Now, what would I do? I would typically go to the account executive and say, "Tell me about the customer. Tell me about the projects. How are they going? What do we know?
“诸如此类。”但我刚刚问了 Claude。于是我去了客户频道,在那里提了个问题,马上就拿到了所有需要的信息,也完全上手了。And so on and so forth." But I just asked Claude. So, I went to the customer channel, and in that channel ask the question, and immediately I had all the required information, and I was fully ramped up.
所以,你可以看到这里发生的事情:我们向 Claude 提出一个问题,它就会作出回应。它是一个助手,对吧?不过,它给出的答案都紧扣公司的背景、公司的知识以及其中的所有信息。So, you can see here what's happening is we ask Claude a question, and it responds. It's an assistant, right? Except that the answers are very much grounded in the company context, the company knowledge, all the information there.
现在这里有很多人。我知道你们都很先进。这不是你们目前的水平,你们其实已经进入了下一阶段。不过,我爸爸确实是这么用的。A lot of people are here right now. Now, I know you all are very advanced. This is not where you are. You've actually moved on to the next step. But, my dad certainly uses it like this.
从任务到职责From tasks to responsibilities
现在,让我们回到 FDE 团队成立初期。当时,客户开始请我们帮忙实施智能体。我们面向客户的应用 AI 团队里,有几个人提出:“也许我们应该考虑协助客户进行实施。”FDE 团队就是这样成立的。Now, let's go back to the early days of the FDE team. It started with customers asking us for help implementing agents. And our applied AI team that works with customers, there were a few people there that said, "Maybe we should think about helping with implementations." And that's how the FDE team started.
一位负责人和几位工程师聚到了一起。Ryan 是我们的负责人,也是他组建了这个团队。他不是真名,但故事是真实的。就像我说的,我们在 Anthropic 经常使用 Slack,所以他为客户合作事宜建了一个频道,几位客户经理也知道这个频道。One leader, couple of engineers, we got together. Ryan was our leader, who started the team. Not his real name, the story is real. And like I said, we use Slack quite a bit at Anthropic, and so he started a channel for engagements, and a few account executives knew about it.
他们开始在那里引入机会并发布信息。They started bringing in opportunities and posting there.
随着这些机会不断涌入,所有这些事情的组织协调都靠 Ryan 的大脑。那就是操作系统。上下文窗口惊人,API 糟糕透顶。所以,机会不断涌来,对吧?Ryan 会说:“好,收到,确认。”As these would come in, the way the organizational glue for all of this was Ryan's brain. That was the operating system. Amazing context window. Terrible API. So, opportunities would come in, right? Ryan would say, "Yes, got it, acknowledged.
我们会想办法解决的。”然后就这样了。需求足够旺盛,没过几周,Ryan 就忙不过来了。那他怎么办?猜猜看?直接问 Claude。于是他就去问了,然后说:“我需要一个跟踪工具。除此之外,再给我一个仪表板,这样我就可以分享给更大的团队。We'll figure it out." And that was it. There was enough demand for this that within a few weeks Ryan couldn't keep up. So what did he do? Any guesses? Just ask Claude. So he went and then he said, "I need a tracker." And along with the tracker, give me a dashboard that I can go share with my broader team.
现在我们有了一个跟踪器和一个仪表板,接下来就可以大展拳脚了。你可以看到,这里发生了这样的演变:Claude 从一个只负责问答的助手,发展到现在由 Ryan 给它布置任务。And now we had a tracker and a dashboard. We're off to the races. So you can see what happened here is the evolution from Claude being an assistant, just question and answers, to now Ryan gave it a task.
而任务就是给我做个追踪器。太好了。And the task is make me a tracker. Great.
但有一个问题:机会不断涌入,追踪表也得随之更新,而这件事必须由 Ryan 亲自完成,或者去请 Claude 帮忙。很快,他就成了瓶颈。而且,这部分工作也并不是他特别喜欢做的。But there was one problem. More opportunities came in and the tracker had to be updated. And Ryan had to do it himself. Or go ask Claude to do it. Very quickly he became the bottleneck. And this was also part of the job that he didn't necessarily enjoy.
于是他向 Claude 提出了请求。这一次,他去了 engagements 频道,把 Claude 添加到频道中,然后说:“Claude,你能把所有商机的信息都保持最新吗?一有新商机进来,Claude 就从 CRM 获取所有数据并填入信息,提出需要澄清的问题;等信息全部补齐后,再去更新跟踪表,随后更新电子表格和仪表板。”So he asked Claude. This time he went to the engagements channel, added Claude to it, and said, "Claude, can you keep all the opportunities updated? As soon as an opportunity comes in Claude gets all the data from the CRM, populates it, asks clarifying questions, and once it's fully populated, goes and updates the tracker and subsequently the spreadsheet, the dashboard.
那我们就可以开始了。So we're good to go.
适用于不断壮大团队的系统A system for the growing team
这里发生的变化是,双方的互动模式变了:从瑞安给 Claude 安排一项任务,变成了把一份责任交给它。要忠于事实。事情就是这样。而第一次真正的交接,就发生在这里。What happened here is this changed the dynamics. It went from Ryan giving Claude a task to giving it a responsibility. Keep this true. That's what happened. And this is where the first real handoff happened.
在人类交给 Claude 一项任务,让它自主完成的场景下。现在,电子表格运行得不错,对吧?机会越来越多,一切都很好。更多 FDE 正在加入,我们开展的合作项目也越来越多。Where the human gave Claude a thing to do autonomously. Now, the spreadsheet works well, right? More opportunities are coming in. Everything is great. And more FDEs are joining. We got more engagements happening.
更多领导者加入进来。团队在壮大,业务也在增长。我们很喜欢这样,太棒了。我差不多也是在同一时间加入的。就像几十年来产品负责人一直对功能完善的电子表格所做的那样,我也做了完全一样的事。More leaders join. The team is growing. The business is growing. We love it. It's awesome. I joined around the same time. Just like product leaders have been doing for decades to perfectly functional spreadsheets, I did exactly that.
转换成了一个应用。我们要开发软件来取代电子表格。话说回来,我这么做也是情有可原:电子表格已经不够用了。Converted into an app. We're going to build software to replace the spreadsheet. Now, in my defense, the spreadsheet wasn't going to cut it anymore.
跨职能团队参与了大量项目,项目处于不同阶段,活动也很多。我们确实需要一个工具,而且需要一个供跨职能团队共同使用的统一工具,作为唯一可信的信息来源。Cross-functional team, lots of engagements, lots of activity, different stages of these engagements. We really needed a tool. And we needed a single tool for the cross-functional team. One source of truth.
那么,我们该怎么做?我们就去开一个 Slack 频道,说我们要做一个跟踪工具。然后在这个跟踪工具里,我们开始反复完善需求,也就是我们想要的规格说明,并且和 Claude 一起推进这件事。So, what do we do? We go and start a Slack channel. And we say we're going to go build a tracker. And in that tracker, we start to iterate on our requirements, the spec that we want, and we're working with Claude on this.
市场推广团队会反馈他们想要什么。在我们不断迭代的过程中,最后我们会说:“很好,那就开始做吧。”对吧?然后 Claude 就会去构建这些东西、发起 PR,然后我们再进行审批。The go-to-market team is giving their feedback on what they want. And as we iterate through it, at the end of it, we say, "Great. Let's go build it." Right? So, then Claude goes and builds these things, opens up PRs, and we then approve.
Ryan 批准了这些 PR。现在我们有了一个跟踪表。应用已经上线。这是一个内部工具。我们已向所有与我们对接的市场推广团队成员开放了这个工具。Ryan approves these PRs. And now we have a tracker. The app was shipped. It's an internal tool. We opened it to all our go-to-market counterparts.
我们向所有用户开放了它。关于用户,有件有趣的事:他们会发现问题,会提出需求,会希望某些东西稍有不同,也会提出疑问。为什么这个不管用?We opened it to all the users. Now, here's the funny thing about users. They find issues. And they have requests. And they want things to be a little different. And they have questions. Why does this not work?
那我该怎么做?诸如此类的问题。于是,我们又建了一个专门收集所有反馈的频道。这次我们告诉 Claude:‘随着反馈不断进来,把问题分类处理,并修复那些显而易见的 bug。’And how do I do that? And so on and so forth. So, we started a channel again for all the feedback. And this time we told Claude, "As the feedback comes in, triage the things and fix whatever bugs are obvious.
“那就去把它们修好。至于新功能嘛,我们会设计出来、制定好规格,然后决定要开发什么。”于是事情就开始这样发展了。你可以看到,用户在这里提交了反馈和投诉。Just go fix them. And whatever the new features are, you know, we'll design them, we'll spec them out, and we'll decide what to go build." And so that's what started to happen. So, here you can see users are filing that feedback, their complaints.
Claude 会逐条回复这些内容,对吧?然后它会去修复这些问题,还会创建这些 PR。这一切都是自动进行的。And Claude is going and replying to each of those, right? And it goes and fixes these. It opens these PRs. Just autonomously this is happening.
这种情况持续下去。此时,Claude 已经实现了工程师的梦想。我们负责构建东西,它负责修复 bug,对吧?但别忘了,决定要构建什么的仍然是我们。大家都很开心。这太棒了。This kept going. At this point Claude has accomplished an engineer's dream. We build things and it fixes the bugs, right? But keep in mind, we still decide what to build. Everyone's happy. It's great.
注意这里发生了什么。我们之前给了 Claude 一项职责,对吧?现在我们又给了它一套要运行的系统。这不只是一项职责,它现在要运行的是一整套系统。这些工具基本上是在自动运行的状态下构建出来的。Note what happened here. We had earlier given Claude a responsibility, right? And now we've given it a system to run. It's not just a responsibility. It has a full system to run. The tooling was built relatively on autopilot.
我们偶尔需要做一些工作。要记住,我们没有专门负责这件事的产品经理和工程团队。所有这些工作都是 FDE 团队在和客户合作的同时,借助 Claude 完成的。There's a little bit of work that we have to do every now and then. Now, keep in mind, we don't have a product manager and an eng team for this. This is just the FDE team that is working with customers and Claude on the side doing all this work for us.
我们遇到了下一个问题。随着工具方面的问题得到解决,瓶颈也随之转移了,对吧?瓶颈已经不再是工具了。We ran into the next problem. As the tooling problem got solved, the bottleneck moved. Right? The bottleneck wasn't the tooling anymore.
人员配置与主动协调Staffing and proactive coordination
猜猜瓶颈转移到哪儿了?转到我们这儿了。我举个例子。就拿人员配置来说吧。人员配置是怎么运作的?周一早上,我们会开项目评审会和人员配置会议。大家聚到一起,所有负责人都聚到一起,而我们分布在世界各地。Any guesses on what the bottleneck moved to? To us. I'll give you an example. Let's take staffing. How does staffing work? Monday morning, we have a deal review and staffing meetings. We get together, all the leads get together, and we're dispersed all over the world.
我们聚在一起,讨论正在接洽的所有项目、已经签下的项目,并开始规划人员配置。我们规模还小的时候,这件事相对容易,对吧?相当简单。We get together, we talk about all the deals that are coming in, the ones that we've signed, start planning staffing. When we were small, this was a relatively easy undertaking. Right? Pretty straightforward.
我们知道谁在做什么,他们的兴趣是什么。完成这个特定项目后,他们是否需要休息一下,还是想继续做下去?他们是否想从事医疗保健领域,诸如此类。We knew who's working on what, what are their interests. Do they need a break after this particular project, or do they want to continue going? Do they want to work in health care, and so on and so forth.
但随着团队不断壮大,我们多少有些不了解当时发生的所有事情。因此,周一早会的前20到25分钟通常都用来讨论当前状况、在不同负责人之间核对信息,并确保大家看到的是同一份数据。But as the team grew, we kind of didn't have the context about everything that was going on. So the first 20 25 minutes of a Monday morning meeting would just go into talking about the current state and reconciling across the different leads, and making sure that we had the same data that we were looking at.
有一次开会时,我们发现 FDE 有一名员工,100% 投入一个客户项目,50% 投入另一个项目,25% 做售前。稍微算一下就能看出,这里的数字有点可疑。好。One of these meetings, we found that there was an individual in FDE who was staffed 100% on a customer project, 50% on another project, 25% doing pre-sales. It didn't need much judgment to figure out the math is a little suspicious here. Okay.
所以你可以看出,缺少的只是信息流通和组织协调。通常会由某种项目经理来替你处理这些事情。So you can see that this is just information flow and organizational glue that's missing. And typically you'd have a project manager of some sorts do this for you.
所以我们又给 Claude 安排了一些工作,对吧?我们对 Claude 说:“每周五,联系所有潜在客户。你去查查追踪表里的交易情况,看看所有备注,联系所有潜在客户,并询问最新进展。”So we gave Claude some more work to do. Right? We said to Claude, "Every Friday, ping all the leads. You should go figure out what the deals are in the in the tracker, look at all the notes, ping all the leads, and ask for updates.
所以,每周五我们都会在交易频道的一个讨论串里收到提醒,Claude 会标记我们并说:“请提供意见。”现在,它不再只是根据某一个信息来源来征求意见,而是会查看与该客户相关的所有 Slack 频道、通话记录和邮件,然后汇总所有信息,再告诉我们:“以下是尚待解答的问题。”然后我们所有人都会回复,对吧?我们会澄清信息,诸如此类。So, every Friday we would get pinged in a thread in our deal channel, and Claude would tag us and say, "Please give inputs." Now, it's not just giving inputs based on what one source of information. It'll go look at all the different Slack channels that are related to that particular customer, call notes, emails, and then pull together all of that, and then come in and say, "Here's the open questions." And then all of us respond, right? We clarify information, and so on and so forth.
Claude 能解决的事情,它都会解决。没有什么比 AI 追着你更能促使你更新状态了。And whatever Claude can resolve, it resolves. There's nothing motivates status updates like AI chasing you.
对吧?效果非常好。所以现在,Claude 会尽可能解决所有问题,而它无法解决的问题,则会留到周一早晨会议的议程上。因此,会议开始前,我们就能拿到议程。Right? It worked very well. So, now Claude tries to resolve everything that it can, and whatever it cannot resolve, it leaves it for the Monday morning meeting's agenda. And so, before the meeting starts, we get the agenda.
好,我们很清楚谁需要讨论什么。会议照常进行,通话会被录下来。然后 Claude 查看文字记录,从中提取所有行动事项。接着我们再审核这些行动事项。All right, we know exactly who needs to talk about what. The meeting happens, the calls are recorded. Claude then goes and looks at the transcripts, and from that extracts all the action items. And then we go and review the action items.
我们会确认需要做些什么,或者确认这些内容是否准确。一旦确认下来,就能帮助我们落实其中的一些行动事项,比如在不同渠道发布消息、向相关人员同步进展、更新跟踪表等等。We confirm what needs to be done, or if they're accurate. And once we confirm, it actually helps us execute on some of these action items. Posting in different channels, updating people, updating the tracker, and so on and so forth.
所以,注意这里发生了什么。如果稍微拉远一点看,注意这里发生了什么。一开始,如果你还记得的话,Ryan 在向 Claude 提问。So, note what happened here. If you zoom out a little bit, note what happened here. In the beginning, if you remember, Ryan was asking Claude questions.
现在轮到 Claude 向我们提问了,对吧?局势反转了。这是交接方式的进一步升级。第一步是让电子表格保持更新,而这次交接又上了一个台阶。这里真正发生的是,Claude 成了我们的队友。And now Claude asks us questions. Right? The tables turned. And this was the next level of handoff. The first one was keeping the spreadsheet updated. This was the next level of handoff. What really happened here is that Claude became a teammate.
如今,Claude 对我们周一会议的了解,比任何一位负责人都多。Claude 承担着跨职能的工作,也负责把各个职能衔接起来。你看沿途的每一步:做个电子表格、建个跟踪表、持续更新、修复 bug、准备会议等等,这些都是渐进式的小事。Today, Claude has more context about our Monday meetings than any of the individual leads. The cross-functional work and the cross-functional glue is Claude. And if you actually look at each of these steps along the way, build a spreadsheet, build a tracker, keep it updated, fix bugs, you know, prep for meetings, they're kind of incremental.
它们是在前一个基础上的渐进式改进。但当你拉远视角,就会发现这其实是一次相当大的转变。它是一个与你并肩工作、融入你工作场景的队友。我们六个月前并没有坐下来就说:“我们要构建一个自主多智能体系统或操作系统。”They're incremental over the previous one. But when you zoom out, you actually see it's a pretty big shift. It's a teammate that works with you where you work. Now, we didn't sit down 6 months ago and say, "We're going to build an autonomous multi-agent or operating system."
谢天谢地,我们没那么做。我们只是不断给 Claude 安排更多工作,并想办法让它发挥作用。Thank god we didn't do that. We just kept giving Claude more work and finding a way to enable it.
AI 团队伙伴的五大构建模块Five building blocks for an AI teammate
那么,我们如何启用 Claude?这里有五个构建模块。第一个是共享上下文,它包含两个部分:第一,我们所有人都可以使用同一个 Claude;第二,Claude 可以访问我们能访问的所有信息。So, how do we enable Claude? There's five building blocks. The first one, shared context. And there's two components to this. One, all of us have access to the same Claude. And then Claude has access to all the information that we have access to.
这一点非常重要,因为它就是通过这种方式了解所有必需的信息:访问工具。我们工作所需访问的所有工具,它也都能访问。And this is very important, because this is how it knows everything it needs to know. Access to tools. All the tools that we need to access we need to get access to do our jobs, it has access to as well.
它有两种访问权限:一种是以 Claude 的身份访问,另一种是以我的身份访问。所以,当我说“去做这个更改”时,它就能代表我进行更改。而且所有操作都会被记录下来,并留有审计追踪。And it has access in two ways. It has access as Claude, and it has access as me. So, when I say go make this change, it can make the change on my behalf. And all the actions are recorded. There's an audit trail.
记忆。我更喜欢事情按照某种方式进行。Memory. I prefer things to be a certain way.
我的同事可能更喜欢稍微不同的方式。我喜欢报告按特定方式组织好后通过电子邮件发送,但他们喜欢在 Slack 上查看。这属于个人记忆。然后还有团队记忆。My peer might prefer them to be a little differently. I like my reports to be structured in a certain way and emailed. But they like it in Slack. So that's individual memory. Then there's a team memory.
然后是组织记忆。你很快就会明白它为什么这么重要。而且它是主动式的,可以通过我调用它来触发,也可以根据时间触发,还可以由某些操作或事件触发。Then there's the organizational memory. And you'll see why this becomes important very quickly. And it's proactive. It can be triggered based on me calling it. It can be triggered based on time. It can be triggered based on certain actions, certain events.
最后一个组成部分是反馈循环。当事情进展不顺利时,我们会给它反馈。它会将反馈存入记忆,进而修改自己的指令。这个循环就这样持续下去。正是这些基础构件,让我们能够在任何工作场景中拥有一位由 Claude 驱动的队友。And the last piece is the feedback loop. When things don't work well, we give it feedback. It puts it into memory. It goes and changes its instructions. And the cycle continues. These are the building blocks that have enabled us to have a teammate powered by Claude wherever we work.
这就是你从助手转变为队友的过程。你知道,大概四五天前,我受邀来这里演讲。And that's how you can go from assistant to teammate. You know, I got invited to come and speak here about 4 days ago, 5 days ago.
临时抱佛脚,对吧?所以我去找了一群同事,说:“各位,我时间不多了。能不能帮我想想我该讲些什么?”直接问 Claude 就行。这事儿差不多就是这么开始的。Last minute, right? So I went to a bunch of my colleagues and said, "Guys, I don't have much time. Can you help me think about what I should go talk about?" Just ask Claude. And that's kind of how this started.
是的。谢谢。Yeah. Thank you.