← 返回任务列表

How to Build Team Agents

404 段 · 2 位说话人 · 原片 41:36
M1
M10:00

2026年可以说是智能体的一年。

2026 has been the year of agents.

从年初的OpenCLAW,到现在的Muse、GrokBot和Instinct等平台,这些平台让人们真正开始利用这些极其强大的自主工具,帮他们完成越来越多的工作。我们确实已经从智能体是下一件大事,走到了它们就在眼前的阶段。

From OpenCLAW at the beginning of the year to now platforms like Muse and GrokBot and Instinct that are getting people to actually take advantage of these incredibly powerful autonomous tools that are getting increasingly large portions of their work done for them, we really have gone from agents being the next big thing to just being here.

问题在于,我们的工作并非只是独立完成的。

The problem is our work isn't just done alone.

我们通常会和别人一起团队合作。

We tend to work in teams with other people.

然而到目前为止,大多数智能体都是单打独斗的,只覆盖我们独立完成的那部分工作。

And yet up till now, most agents have been solo affairs, only covering the portion of our work that we do on our own.

我认为现在这种情况正在转变,我把这种趋势称为多人AI,或者共享智能体、团队智能体。

I think that is shifting now, a trend which I've talked about as multiplayer AI or shared or team agents.

但到底什么是构建团队智能体呢?

But what does it mean to even build a team agent?

需要考虑哪些因素?

What are the types of considerations that go into it?

它与仅仅为自己构建一个智能体到底有多大不同?

And how different is it really than just building an agent for yourself?

这些问题正是我在AI Daily Brief的Operator's Cut版中与Noufar Gaspar探讨的内容。

Those are the questions that I get into with Noufar Gaspar on this Operator's Cut edition of the AI Daily Brief.

《AI每日简报》是一个每日更新的播客和视频节目,关注AI领域最重要的新闻和讨论。

The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

好了朋友们,在正式开始之前,先快速宣布几件事。

All right, friends, quick announcements before we dive in.
M1
M11:27

进入正题前还有几点要提一下。

Just a couple other notes before we get in.

显然,这是一期预先录制的节目。

Obviously, this is a prerecorded episode.

今天有不少事情正在酝酿。

There are a bunch of things cooking today.

我们会有很多要聊的,所以明天我们会恢复常规形式回来。

We will have a lot to talk about, so we will be back with our normal format tomorrow.

我还想分享一些即将到来的机会。

I also wanted to share a couple of upcoming opportunities.
M1
M12:41

好了,闲话少说,让我们来谈谈如何构建团队智能体。

All right, with all that out of the way, let's talk about how to build team agents.

好的,Nufar,欢迎回到节目。

All right, Nufar, welcome back to the show.

我今天准备了一期Operator's Cut。

We got an operator's cut today.
F1
F12:52

是的。

Yes.

很高兴再次来到这里。

Happy to be here again.
M1
M12:54

这期的起源是我们在入秋前的一些对话,关于我们想在这个秋季的免费自主学习训练项目中做些什么。

This one has its genesis in some conversations we were having as we were coming up into the fall around what we wanted to do with the, you know, this fall's edition of a free self-directed training program.

我们当时聊了很多关于多人AI的概念,以及一种正在转变的模式:从人们只为自己构建单独使用的智能体,转向我们预测将要开始看到——我想我们已经开始看到早期迹象了——更多存在于共享工作空间中的智能体。

And we were talking a lot about this idea of multiplayer AI and a shifting pattern from people just building solo agents that they were using themselves to a prediction that we're gonna start to see, and I guess we're starting to see early evidence of, more agents that live in between people's shared workspace.

这其实只是遵循了人们自然的工作方式。

And this kind of just follows the natural way that people work.

你的很多工作是独立完成的,但也有大量的工作是在与其他人交汇的地方完成的,而团队智能体就可以存在于这些地方。

A lot of your work is done individually, but then lots and lots is also done at the intersection with other people, and that's where team agents can live.

当我们正在构建那门课程(目前可以在Multiplayer AI上获取)并更广泛地思考这个概念时,我们反复提到的一个点是,这个领域还太新了,以至于实际构建一个团队智能体意味着什么并不一定非常明显。

And as we were building out that course that's available right now at Multiplayer AI and just thinking about this concept more broadly, one of the things that we kept coming back to was that this is nascent enough that what it means to actually build a team agent won't necessarily be super obvious.

所以今天这期Operator's Cut的目标是帮助大家真正思考如何构建团队智能体,理解团队智能体是什么样子的,了解它们与人们可能已经构建的智能体类型有何不同——或者我认为我们在这里会论证的是它们其实很相似——以及人们接下来可以往哪个方向发展。

And so the goal of today's Operator's Cut is to help actually think through how to build team agents, to understand what team agents look like, to understand in what ways they are different from, or I think probably what we'll argue here, similar to the types of agents that people might have already built and where they can go from here.

所以非常高兴你能回来,也很兴奋能深入探讨这个问题。

So super excited to have you back and excited to dive in here.
F1
F14:12

太棒了。

Amazing.

所以我要拓宽你的定义,把它叫做团队智能体。

So I'm going to broaden your definition and I'm going to call them team agents.

这个概念就是,你的整个团队都能用的智能体,要么是因为它们之间协作完成工作,要么就是它们可以在团队成员之间共享。

And the concept is teams that your entire team can work with, whether it's because work happened between them or just because they are something that can be shared across team members.

简单来说,有些智能体应该保持个人私用,而另一些应该变成团队级别的智能体。

And kind of the short version is that some agents should stay yours and private while others should become the team-level agents.

那些成为团队智能体的,需要有意做出几个决策。其中一些决策,就像你说的,跟任何好的智能体配置都一样,另外一些则更独特,或者至少需要更刻意地考虑。

And the ones that do become team agents need a few decisions made on purpose. Some of them, as you said, overlap with any good agent configuration, and some of them are more unique or at least more intentional.

这就是我们要讨论的内容。

And that's what we'll walk through.

我想先通过两个故事来激励你们,这两个故事你们在自己的公司或生态里很可能也见过。

And I wanted to start as a means of motivation to give you like 2 stories that you will probably recognize from your company or your ecosystem.

第一个故事说的是一个人——跟我合作过的每家公司、每个人身边至少都有这么一号人。

So the first is about a person that everybody that I work with, every company that I work with has at least one like that.

这个人非常清楚定价例外是怎么运作的,数据是怎么回事,或者我们最大的客户三年前是怎么配置的。

And this person, they really know how their pricing exception works or what the data is all about or how our biggest customer setup was configured 3 years ago.

当ta忙得不可开交时,工作就得等着ta,因为只有ta知道。

And when they're swamped, then work has to wait for them because they're the only one who knows.

ta休假的时候,还是会有人给ta打电话。

And when they're on vacation, someone still calls them.

ta一旦离职,公司的一部分也跟着走了。

And when they leave, a piece of the company leaves with them.

在我合作过的一家公司里,每个领域都有这么一个角色。

So in one of the companies that I work with, they had, I think, a person like that for each and every domain.

所以没人能休假而不被同事打电话打扰。

So no one gets to take vacation without getting a call from their peers.

显然,这不是理想的状态。

And obviously that's not a desired state.

第二个场景是没人完全负责的工作。

The second scenario is work that nobody fully owns.

比如一个客户从销售转到市场再转到客户成功,销售在谈交易时做了承诺,市场又在用稍微不同的信息做活动,然后客户成功在续约前几周才发现这些承诺,每个团队都有自己那一块,但没有一个人掌握全局,因为工作分散在各个团队之间。

So a customer can move from sales to marketing to customer success, and then sales made them a promise during the deal conversation, and marketing is running a campaign with slightly different messaging, and then customer success finds out these promises were made to them a few weeks before the renewal, and each team has their own piece and nobody has the whole picture because the work sits between them.

就像你说的那样。

So to your point.

就算每个环节都用了AI,这些智能体之间也不沟通,很多时候反而让问题更严重。

And even if they are using AI in each step of the process, these agents don't talk to each other and only worsen the problem in many cases.

所以一个为整个团队打造的智能体可以帮我们解决这两个问题——当然,这两个问题性质很不一样。

So an agent that is built for the whole team can help with both of these problems, and they are, of course, quite different problems.

还有一个原因我觉得现在这件事很重要,你应该关注起来,哪怕你觉得这有点超出你的理解范围——那就是我们在最前沿的AI公司中开始看到的模式。

So there is another reason I think why this matters right now and why you should pay attention now, even if you feel a little bit like this is above your head, and that's the pattern that I think what we're starting to see across the most AI-forward companies.

这个模式大致分三步。

And it goes in basically 3 steps.

第一步,每个人都建自己的智能体,生产力提高了,大家都很开心。

The step one is that everybody builds their own agents and they're happy with their productivity boost.

但如果跑着太多这样的智能体,就会出现智能体泛滥的情况。

Only with enough of those running around, we kind of get into an agent sprawl.

很多智能体在做重复的工作,每个都由一个人维护,每个对公司情况的了解都略有不同,而且一旦维护者失去兴趣或离职,这个智能体就报废了。

Lots of agents doing overlapping work, each maintained by one person, each with slightly different picture of the company, and each one stops being useful the day that the owner loses interest or leaves the company.

然后,在最前沿的AI公司里,你会看到他们开始把一些智能体合并成团队级别的智能体。

And then what you see in the most AI-forward companies they started to merge some of those agents into team-level agents.

这些智能体通常数量少得多,范围广得多,每个都有明确的负责人,被很多人使用,并且随着团队不断了解它们该做什么不该做什么,会逐步优化。

Those will typically be much fewer agents, much broader in their scope, and each with a named owner and used by many people and ideally refined over time as the team learns what they should and shouldn't do.

这种事情正在非常公开地发生。

And this is happening very publicly.

已经有很多公司在讨论了。

There are many companies already talking about it.

我记得你提到过Everys的经验。

I think you mentioned Everys' experience.

他们年初的时候给每个员工都配了一个智能体,然后到了五月份,已经转向了共享的团队智能体。

They started by giving every employee an agent early in the year, and then by May they have moved to shared team agents.

Sierra 把他们很多专门化的 agent 合并成了一个。

Sierra merged many of their agents specialized into one.

Shopify 的内部 agent。

Shopify's internal agent.

所以我们看到很多这样的例子,出现在各地非常公开的、面向公众的公司里。

So we see a lot of these in very public speaking companies all over the place.

而我看到的大多数团队可能都还处在第一步或第二步,但我认为我们所有人都很重要的一点是去看这些 AI 领先的公司,理解怎么到达第三步,以及怎么正确地做。

And most teams that I see are probably either in step 1 or 2, but I think that it's very important for all of us to look at these AI-forward companies and understand how to get to number 3 and how to do it properly.

这就是今天整期的目的。

And that's the entire purpose of today.

为了确保我们说的是同一件事,因为基本上同一个东西有好多不同的名字。

So to make sure that we are talking about the same thing, because there are multiple names to basically the same thing.

有些人,包括你自己,叫它 multiplayer AI。

Some people, including yourself, call it multiplayer AI.

你可能也听过 shared agents。

You probably also heard shared agents.

有些人称它们为 AI teammates,甚至有时候还会把 company brain 也混进来,或者说跟它混用,都是指在公司范围内使用共享知识的 agent。

Some people refer to them as AI teammates, and even company brain is sometimes thrown into the mix or interchangeably used to mean agents being used with shared knowledge across the company.

在这个节目里,我会一直称它们为 team agents。

I'm going to refer to them throughout the episode as team agents.

我的意思是,我们有一个 agent,很多人跟它对话。

And what I mean by that is we have one agent that many people talk to.

共享的知识、共享的记忆,还有一个单独的配置。

With shared knowledge, shared memory, and one configuration.

指令、技能、访问权限,还有所有者,全都是共享的。

The instructions, the skills, the access, and the owner, they are all shared.

你可能在听到我这么说的时候想,我们不是已经共享 skills 了吗,对吧?

And you might be thinking when you hear me saying that we already share skills, right?

大多数公司都有一个很棒的 skill library,或者正在搭建 skill library,这很棒。

Most companies have an amazing skill library or working on a skill library, and that's awesome.

这是你们在公司里做工作的一种很好的标准化。

And that's great standardization of how you do the work in the company.

但一个 skill 是针对特定任务的 playbook,而 team agent 是你的整个团队用来处理各种不同任务的。

But a skill is a playbook for a specific task, where a team agent is something that your whole team works with on diverse set of tasks.

既有临时性的,也有重复性的工作。

Ad hoc as well as repeated stuff.

它确实携带团队知识,记住学到的东西,并且使用团队级别的 skills,如果你们有的话。

It does carry the team knowledge, remembers what it learns, and using the team-level skills if you have them.

当然,那些也可以接入 skills 市场等等。

Those, of course, can also tap into skills marketplaces and so on.

所以 skill library 也许是其中的一个要素,但它们是两码事。

So a skill library perhaps is one of the ingredients, but they are not one and the same.

而我希望你今天思考的是,如何以及何时开始构建你的下一个 team agent。

And I want you to today think about how and when to start building your next team agent.

关于范围再补充一点,因为现在大家对各种 personal agents 非常兴奋。

And one more note on scope, because there's a lot of excitement right now about all of the personal agents.

我说的是 Muse、Instinct,还有其他一些这个类别里的东西。

I'm talking about Muse and Instinct and some of the others in this category.

那些是用在家里或者私人生活中的。

Those are for like a home or private life.

今天我们的重点是工作。

Today the focus is going to be on work.

所以这是第一件事,确保范围清楚。

So that's one thing to make sure that it's clear about the scope.

我们说的是你为工作构建的 agents。

We're talking about agents that you build for your job.
F1
F112:44

好的。

All right.

我想确保大家明白这期节目是给谁做的。我觉得它是为所有人做的,不只是那些在搞最尖端东西的前沿专业人士。

I want to make sure that we understand like who I built the episode for, and I think it's built for everybody and not just the frontier professionals that are building the absolute cutting edge.

如果你正要构建一个,当然要留意,因为它会为你提供或验证完整的操作手册。

If you're about to build one, of course, pay attention because it will provide you or verify the full playbook.

它也是针对那些尚未完全掌握的人,因为正如你所说,这些决策确实适用于你构建或使用的任何智能体,而团队智能体只是让其中一些决策变得更加关键。

It's also aimed at people who are not quite there yet because the decisions, as you rightfully said, do apply to any agent that you build or use, and a team agent just makes some of them even more critical.

而且即使你是独自工作,并且你有一组智能体或者正在考虑构建一组智能体,你也会面临同样的决策。

And even if you are working solo and you have a team of agents or you're contemplating building a team of agents, you have the same decisions.

你生态系统中的其他参与者可能不是你的同行,因为你独自工作,但也许你是在为你的客户构建它,或者为未来的自己构建,以确保它足够稳健,并且能充分代表多样化的工作任务。

The other players in your ecosystems are probably not your peers because you work alone, but perhaps you're building it for your customers or you're building it for a future you to make sure that it's robust enough and representative enough of a diverse set of work.

所以这就是哪些人需要关注的原因。

So that's the motivation for who should pay attention.

我想确保一点非常清楚,那就是并非每个智能体都应该被共享。

And one thing that I wanted to make sure that it's very clear is that not every agent should be shared.

这里有一个拨盘或者说一个频谱,有三个档位。

There is a dial here or a spectrum with 3 settings.

我们有一个私有智能体,它属于你,用于你的工作,体现你的品味和你的权限。

We have a private agent that's yours for your work with your taste and your access.

例如,我自己的社交媒体智能体就是私有的。

For example, my own social media agent stays private.

我可能无法与任何人分享它,因为我是唯一一个想以特定方式在社交媒体上分享或发帖的人。

I'm probably not gonna be able to share it with anybody because I'm the only one that wants to share or write on social in a specific way.

所以没有人应该像我一样说话。

So nobody should ever sound like me.

然后我们有共享知识。

And then we have shared knowledge.

这些可以是共享的知识和技能,但仍然是私有智能体,意味着团队维护一个知识体系。

Those can be shared knowledge and skills, but still private agents, meaning the team maintains one body of knowledge.

例如,我们销售什么,我们如何工作,我们的用语是什么意思,通常还有一个共享技能库,然后每个人将自己的智能体指向那个库。

For example, what we sell, how we work, what our words mean, often with a shared skill library, and everybody points their own agents at that.

但顺便说一句,这就是技能库所在的地方。

But this is where the skill library lives, by the way.

而且这通常是正确的答案,也是最容易的起点。

And it's often the right answer, and it's the easiest place to start.

另一个具体例子,比如每个销售员都有一个根据自己风格和偏好调整的寻找客户智能体。

Another concrete example, say every salesperson has a prospecting agent tuned to their own style and their own preferences.

保持这些智能体的原样,因为每位销售员都想拥有自己的声音,但给他们所有人都提供相同的维护良好的理想客户画像和公司传达的信息。

Keep those agents as they are because every salesperson wants to have their own voice, but give them all the same well-maintained picture of the ideal customer and the messaging for the company.

所以这是一种混合模式,某些公司或某些用例应该保持这种模式。

So that's a hybrid mode that some companies or some use cases should remain.

最后,我们确实有团队智能体。

And lastly, we do have the team agents.

在这种情况下,我们有一个智能体,多个人与之协作,它有自己的工作和自己的所有者。

Where we have one agent that many people work with that has its own job and its own owner.

并且智能体可以在这些档位之间移动。

And agents can move along the dial.

如果你遇到这样的情况,同事不断要求借用你的私有智能体,那很可能表明你需要把它变成团队智能体,或者考虑与他人共享它。

And if you have a scenario where your colleagues keep asking to borrow your private agent, that's probably a sign that you need to make it a team agent or to consider sharing it with others.

所以我想回答的下一个问题是,所有团队智能体都来自同一个原型吗?还是它们都遵循相同类型的用例?

So the next question that I want to answer is, are all team agents from the same archetype, or do they all follow the same type of use cases?

答案是否定的。

And the answer is no.

比如,在我合作的团队中,我大致可以将现有或未来构建的团队智能体分为四种。

Like, across the teams that I work with, I can roughly categorize the existing or future built team agents into 4 kinds.

了解你的类型,不仅有助于你识别用例,还能帮你优化用例。

And knowing which your kind is, it helps you not only identify use cases, but also refine the use cases.

它还能告诉你需要注意什么才能做对。

And also it can tell you what to pay attention to in order to get it right.

所以第一种团队智能体,我称之为专家智能体。

So the first type of team agent, I'm calling it the expert agent.

它是一个人或者一个小团队的专有知识。

It's the one person's know-how or one small team's know-how.

它对所有依赖它的人来说都是可用的。

It's available to everyone who depends on it.

你可以通过那个能休假的人认出它来,还记得开头说的吗?

You'll recognize it by the person who can take the vacation, you remember from the beginning.

如果你想再举个例子,可以是那个能回答公司里多个部门任何数据问题的数据助手,或者是一个服务于很多市场团队的定价和交易台助手,还有合规助手,等等。

If you want another example, it can be the data agent that can answer any data question across multiple departments in the company, or a pricing and deal desk agent that serves a lot of go-to-market organizations, compliance agent, and so on.

要想把它做对,知识必须来自公司里掌握这些知识的专家。

In order to get it right, the knowledge has to come from the experts that hold it in the company.

他们需要被访谈。

They need to be interviewed.

他们需要——你需要收集他们在多个场合已经给出的答案,比如即时消息、邮件和其他地方,而且这些专家必须从一开始就参与进来,因为对他们来说,这最终让休假成为可能,但也带来很多工作不安全感,所以在处理这个领域时要小心。

They need to— you need to collect the answers they already gave in multiple forums, whether those are direct messaging or emails and other places, and they have to be— the experts have to be involved from day one because for them, this is what finally makes the vacation possible, but also a lot of job insecurity, so tread carefully when working in this domain.

第二种类型我称之为通用工作助手。

The second type is what I refer to as the common work agent.

这种情况是很多人都在做类似的重复性工作,并且只有一种共同的方法来做。

That's the scenario where many people are doing similar recurring work with one shared way to do it.

识别这种用例的方法是,当三个人各自建了自己的同一个助手时。

The way to recognize a use case that falls into this category is when 3 people have each built their own version of the same agent.

你可以想一个团队研究和会议准备助手。

You can think about a team research and meeting prep agent.

这是很经典的一个,或者市场团队的内容助手。

That's a very classical one, or marketing team's content agent.

我相信你还能想到其他的。

And I'm sure that you can think of others.

要想把这个类型做对,你得就工作方式达成一致,这说起来容易,做起来难,因为你是在合并三个版本的最佳部分,这实际上是一场关于你在公司标准上达成什么共识的对话。

In order to get this archetype right, you have to agree on how work is done, and it's easier to say than to actually execute because you're merging the best of 3 versions, and that's really a conversation about what you agree in terms of the standard for the company.

所以,一旦你开始考虑统一这样一个助手,这至少会是一场有趣的对话。

So an interesting conversation at the very least once you start contemplating unifying an agent like that.

然后我们有桥梁助手。

And then we have the bridge agent.

这是那种在角色之间流动的工作,没人能独自完成。

That's the work that flows between roles where nobody can do it alone.

当交接出问题时你就能认出来。

You'll recognize it when the handoffs break.

每个阶段都必须解释上下文,或者这些助手必须在不同部门之间协同工作。

And every stage has to explain the context or the agents have to somehow work together between different departments.

例子可以是一个跨越销售方案、客户成功和交付的客户助手。

Example can be a customer agent that spans sales solution, customer success, and delivery.

这是非常经典的一个。

That's a very classical one.

要想把它做对,每个功能都必须有自己的知识片段输入到这个助手里。

And to get it right, we have to have each function their own piece of knowledge that is being fed into this agent.

拥有不同权限的人最终都会用到它。

And people with different access will be eventually using that.

所以我们必须非常非常注意权限问题,并且需要在这方面下大功夫。

So we have to also pay attention very, very carefully to permissions and you'll have to work hard through that.

最后,我们有参谋长助手。

And lastly, we have the chief of staff.

团队运营都由这个助手掌管,比如日常工作的运作执行。

That's the agent that owns the team operating, like operationalizing of the day-to-day work.

它可以处理决策、承诺、状态、入职培训,当团队不断重复自己,新成员需要几周才能站稳脚跟时,你就能认出这个助手。

It can be the decision, the commitment, the status, onboarding, and you will recognize this one when the team keeps repeating itself and new joiners take weeks to find their footing.

要想把它做对,你需要明确界定它允许学习什么,如何学习流程和正在进行的部分,以及如何自动做到这一点,这可不是件简单的事。

And in order to get this one right, what you need to do, you need to clearly define what it is allowed to learn, how can it learn the processes and the ongoing, and how does it do so automatically, which is not very trivial.

当然,还有很多关于权限之类的问题。

And then there are also, of course, many questions around permissions and so on.

所以,当你在思考这些类型,以及它们中哪一个可能适合你正在考虑或应该考虑的一些用例时,在我给你构建团队助手的实操指南之前,先快速绕个弯,因为我想做个简单的现实检查。

So while you're thinking about these archetypes and which one of them might fit some use cases that you are pondering or that you should be thinking about, and before I give you the playbook on how to actually build these team agents, a quick detour because I do want to give a quick reality check.

有些迹象表明,团队助手是个错误的选择,或者至少现在对你不合适。

There are some signs that a team agent is the wrong move, or at least not the right move for you at this moment.

所以一个你不应该构建团队智能体的迹象,至少是现在还不行,就是当个人品味胜过统一标准的时候。

So one indication where you shouldn't build a team agent, at least yet, is when taste beats standards.

如果不同的人真的需要不同的答案,或者不愿意同意一个标准,并且他们自己的判断或他们自己的声音才是重点,那么这些就应该保持为私人智能体,这样人们才能保持真实,不必为基本事实而争论。

If different people truly need different answers or are not willing to agree on a standard and their own judgment or their own voice is the point, those need to remain private agents so people can remain authentic and not have to fight about the ground truth.

第二个不应构建的迹象是,没有人能拥有这些知识。

And the second indication not to build is nobody can own the knowledge.

如果团队不能就工作如何完成达成一致,或者没有人愿意长期拥有和维护共享知识,智能体将在几周内,有时甚至是几天内偏离轨道。

If the team cannot agree on how the work is done or nobody is willing to own and maintain the shared knowledge over time, the agent will drift within weeks, sometimes within days.

所以在你去构建团队智能体之前,先理清所有权问题,因为那会是一个不可行的情况。

So sort out the ownership before you go and build the team agent because that's going to be a no-go.

最后,每当你意识到尝试构建一个共享智能体,一个团队智能体,只是让事情变得更复杂而不是更简单,因为你有相互冲突的需求或纠缠不清的权限,无休止的协调。

And lastly, whenever you're realizing that trying to build a shared agent, a team agent only complicates more than it simplifies because you have conflicting needs or tangled permissions, endless coordination.

如果你意识到结果比智能体能为你提供的价值还要多的工作量,那就是答案。

If you realize that the result is more work than what the agent can provide for you, that's the answer.

不要构建它,至少在你能够理清一些复杂性之前不要。

Don't build it, at least not until you are able to untangle some of the complexities.

当然,注意列表里缺少了什么:敏感数据和高风险。

And of course, notice what's missing from the list: sensitive data and high stakes.

那些是设计问题,它们会塑造你如何构建它,这也是我们接下来要讨论的。

Those are design questions and they shape how you build it, which is where we're going next.

所以我不认为数据过于敏感是反对的理由,它只是需要额外小心的关注。

So I don't think that when data is overly sensitive is a reason against, it's just something that needs extra careful attention.

为了构建候选用例,希望你已通过了之前分享的决策检查清单,这归结为你的智能体的5个核心设计决策。

In order to build the candidate use case that hopefully you've gone through the decision checklist that I shared before, it comes down to 5 core design decisions for your agent.

它们包括它做什么,它在哪里,它知道什么,它能接触什么,以及你如何运行它。

It goes through what it does, where it lives, what it knows, what it can touch, and how can you run it.

这些是核心问题。

These are the core questions.

为了让它更具体,我将从头到尾使用一个例子。

In order to make it more concrete, I'll use one example the whole way throughout.

它将是一个客户智能体,是那种桥梁型的,意味着它持有公司对每个客户所知道的一切。

So it's going to be a customer agent and it's going to be the bridge kind, meaning one that holds everything the company knows about each customer.

以及我们承诺给他们的一切,它可能会被销售、解决方案工程、客户成功交付等使用。

And everything we've promised them, and it's probably going to be used by sales and solution engineering and customer success delivery and so on.

他们将使用那个示例智能体来进行各种与客户相关的活动。

They will be using that example agent to do various customer-related activities.

那么让我们分解其中一些决策,让它变得可操作。

So let's break down some of these decisions to make it actionable.

首先,你必须做出的决策是它做什么。

So first, the decision that you have to make is what it does.

这里有一个快速的提醒,对于任何认真的工作智能体来说,有很多范围界定是非常相似的。

A quick caveat here, there is a lot of scoping that is very similar for any serious agent at work.

它需要有一个清晰的工作描述,一个完成的定义,以及一个它不做什么的列表。

It needs to have a clear job and a definition of done and a list of what it doesn't do.

我在这里将坚持讲,当它是为团队构建时,与为你自己构建的智能体相比,有什么变化。

I'm going to stick here to what changes when it's built for a team versus an agent that you build just for yourself.

所以第一个决策是按角色划分,它为谁服务。

So the first decision is who it serves by role.

例如,销售问它的东西和交付团队不同。

For example, sales asks it different things than delivery does.

所以要写下每个角色以及他们会来找智能体做什么,以确保你覆盖了所有的范围。

So write down each role and what they'll come to the agent for to make sure that you're covering all the scope.

当然,你可以瞄准一个广泛的工作领域。

And of course you can aim for one broad area of work.

因为我认为团队智能体应该是相当有能力的智能体,否则很难证明它们存在的合理性。

Because I think the team agents should be quite capable agents, otherwise it's harder to justify their existence.

在我们的例子中,我预期客户智能体能够准备会议答复,我们与这个客户的关系进展如何,能够标记涉及其他团队的承诺,并且草拟每一次交接。

In our example, I would expect that the customer agent will be able to prepare meeting answers, where do we stand with this customer, will be able to flag promises that commit other teams and draft every handoff.

例如,我有一个相当广泛的范围,让它保持专注的是工作领域。

For example, I have a quite broad scope and what keeps it focused is the area of work.

在我们的情况下,这主要指的是客户。

And that is primarily in our case, customers.

我还想让你注意一下团队禁止清单。

I want you also to pay attention to the team don'ts list.

这部分人们经常容易跳过。

This is the part people often tend to skip.

比如,它绝不代表别人做承诺,也绝不解决人与人之间的分歧。

For example, it never makes commitment on someone's behalf, or it never settles disagreement between people.

这些事都交给它的主人。

Those go to its owner.

你可以想想你们那边还有没有类似的例子。

You can think of another such examples in your case.

还要确保禁止清单里包含权限和数据处理的条目。

Also make sure that the don't list includes permissions and data handling stuff.

举个例子,它绝不会把某个私人空间的信息带到共享空间里。

So for example, it never carries information from one private space into a shared one.

所以它不会在另一位客户的空间里讨论某位客户,也不会代替某个人对另一个人发言。

So it never discusses one customer in another customer space, and it doesn't speak for one person to another.

当然,跟任何 agent 一样,最好从较窄的范围开始,比如读取和起草。

And of course, like with any agent, ideally start with narrower scope, reading and drafting.

只有当它获得了足够的信任并且经过充分验证之后,你才能随着信心的增长逐步扩大范围。

Only when it earns sufficient trust and was validated enough, then you can increase the scope as you gain more and more confidence.

那么这就是第一个决策,接下来我们可以进入下一个。

So that's the first decision, and we can move to the next one.

第二个问题是 agent 存放在哪里,这个问题也是问得最多的。

The second question will be where the agent lives, and this is the question that gets asked most.

所以我们具体一点来说。

So let's be a little bit concrete.

基本上,如果我尽量简化的话,大概有三种分享 agent 的方式,从最简单的到最复杂的。

Basically, if I'm trying to make it as simple as possible, there are roughly 3 ways to share an agent from the simplest to the most involved.

最简单的方法就是用一个共享文件夹,搭配你们公司已有的工具。

The simplest method will just be to create a shared folder with your own tools, like the tools that your company already owns.

然后每个人把自己的工具指向这个共享的 agent,文件夹里会包含指令以及如何存储新信息的方式,作为 agent 整体行为的一部分。

And then each person just points their own tool to the shared agent and the folder will include instructions and potentially how to store new information as part of the way the agent overall behaves.

这是一种非常初级和基础的方式,但作为构建共享 agent 和团队级 agent 的垫脚石,它是一个很好的起点,特别是如果你的团队成员已经在使用类似的 agent 工具,或者可以指向文件夹作为信息来源的 agent 工具。

That's a very naive and basic way, but as a stepping stone to building shared agents and team-level agents, that can be a very good start, especially if all of your team members are already using similar agentic tools or agentic tools that can point to folders as sources for their information.

这就是第一种也是最简单的方法。

So that's the first and simplest method.

第二种方式是使用供应商预制的 agent,我们看到越来越多这样的产品,并且相信在未来几周和几个月里我们会持续看到更多。

The second way that we can do that is using a vendor-ready-made agent, and we're increasingly seeing more and more, and we believe that we will continuously see more and more in the coming weeks and months of the year.

我们后面会详细讲,但它的工作方式是供应商托管它,然后你来配置。

We'll talk more about it later, but the way this works is that the vendor hosts it and you configure it.

这类工具中有很多最近很火的,比如 Claude Tag。

And in this category, we have many very recently famous tools, including Claude Tag.

OpenAI 有自己的 ChatGPT Workspace agent。

OpenAI has their ChatGPT Workspace agent.

Copilot 也有自己的类似产品可以用来做这件事。

Copilot has their own offering that can be used like that.

Notion 和其他很多工具都已经提供了共享空间。

Notion and many others are already offering shared spaces.

你可以跟别人一起在 agent 上协作,只需要配置好它们的细节就行。

With agents that you can work together on, and it's just a matter of you configuring their specifics.

最后,当然也是最复杂的,就是自己托管的 agent,也就是你自己运行的。

And lastly, and that's of course the most sophisticated, is an agent that you host, meaning that's something that you run.

比如,它可以是一个开源 agent,像 OpenCLaw 或 Hermes,放在你自己的服务器上或者租用的云上,有些公司甚至会根据这些需求自己定制 harness。

It can be, for example, an open-source agent like OpenCLaw or Hermes on your own servers or on a leased cloud, or some companies are even building their own custom harnesses specifically for these needs.

这是最复杂的方式,显然技术上的要求也最高,但同时自由度也最大,可以围绕它进行构建。

So that's the most sophisticated, but obviously has the most tech, like technically demanding requirements, as well as the most freedom to build around that.

这就是团队 agent 可以存放的三个大方向。

So that's the 3 broad strokes options of where these team agents can live.

除了怎么构建工具这些决策之外,还有两个随之而来的问题。

On top of the decisions of how to build the tools, there are 2 additional questions that come with them.

第一个问题是,谁能看到每个人和智能体之间的对话?

The first question is who can see each person's conversation with the agent?

也许只有正在和智能体对话的人才能看到。

Maybe it's only the people who are conversing with the agent.

也许是整个频道,或者是会话里的所有人。

Maybe it's the entire channel or everyone in the session.

在这方面,不同工具之间差别很大。

Over here, there are a lot of differences between the different tools.

在某些工具里,每个人都能读到所有聊天记录。

In some tools, everybody can read every chat.

比如在 Slack 里的 Claude,频道里每个人都能看到智能体在做什么、它跟别人聊了什么,而且还能引导它。

For example, in Claude in Slack, everyone in the channel sees what the agent does and what the agent converses with others and can also steer it.

而很多其他智能体,你的聊天是私密的。

With many other agents, your chat is private.

所以这有时是厂商的设计决策。

So that's sometimes a design decision by the vendor.

有时是你可以配置的。

Sometimes it's something that you can configure.

第二个问题是它存储在哪里。

And the second question is where it's stored.

以及谁能读到?

And who can read it?

所以理想的智能体不是那种明显固定死的,而是拥有大量记忆、能随时学习的。

So an ideal agent is not one that is obviously frozen, but one that has a lot of memory and learning on the go.

然后问题是,这些学习内容存储在哪里?又是怎么发生的?

And then the question, where is this learning being stored and how does it happen?

它是按人来发生的,智能体只从和你的互动中进化;还是按频道或团队来?又或者整个工作区有共享的学习和记忆,智能体在公开场合和所有人一起进化?

Is it something that happens per person and then the agent evolves just from its interaction with you, or is it per channel or team, or maybe the entire workspace has a shared learning and memory and the agent evolves with everybody in public?

当然,有一件事,绝不要跨越客户之间的界限。

And of course, one thing, never cross customers.

所以在第一个真实任务之前,一定要检查、选择,并告诉你构建的团队智能体处于什么状态,因为信任的建立或丧失就在于此。

So do check and choose and tell the team before the first real task what's the status with the team agent that you built, because this is where a lot of trust can be gained or lost.

我的原则是选择最简单的方案,只要这周真的有两个人会用就行。

And my rule is to pick the simplest option that 2 people will actually use this week.

对我们的客户智能体来说,那可能是频道智能体,或者一个简单的,比如工具智能体。

And for our customer agent, that's probably a channel agent or a simple, like a tool agent.

因为完整的功能需要它。

Because full functions need it.

如果我把它搞得太复杂,人们就得先弄懂怎么去连接那个东西,而不是直接进入 Slack 或 Teams 频道,那肯定行不通。

And if I will make it overly complicated and people will need to understand how to connect to that versus just going into a Slack or Teams channel, it's not gonna work.

所以以我们的情况,那可能就是正确的解决方案。

So in our case, that's probably gonna be the right solution.

这个决策,我可以说,大概是最重要的决策,那就是它知道什么。

This decision, I wanna move arguably to the most important decision, and that's what it knows.

我觉得如果你构建过任何智能体,这个方法听起来会很熟悉。

And I think if you've built any agent, the recipe will sound very familiar.

对团队来说,不同的是,这正是团队对‘基本事实’达成一致的时刻。

What's different for a team is that this is the moment the team agrees on the ground truth.

工作到底是怎么完成的,我们用哪些定义,哪个版本的价格政策才是真的,而且我觉得这些对话很值得进行,哪怕你永远不发布这个智能体。因为在大多数团队里,哪怕只是就知识达成一致都是一件大事,而且也是团队智能体变难的地方。因为知识一旦共享,就会有更多的贡献者、更多的矛盾,也就会有更多地方可能让重要的事情被遗漏。

How the work actually gets done, which definitions we use, which version of the pricing policy is the real one, and I think that that conversation is worth having even if you never ship the agent because in most teams, even just agreeing on the knowledge is a big deal and it's also where team agents get harder because the moment knowledge is shared, then you have more contributors and more contradictions and more places for something important to fall through.

所以围绕知识的管理过程,比在你的个人智能体里任何时候都重要得多。

So the process around the knowledge matters even more than it ever did in your private agents.

所以你需要非常注意这一点,理想情况下你应该经历这四个阶段。

So you need to pay careful attention here, and ideally you should go through these 4 stages.

我想让你从收集智能体开始,方法是采访掌握知识的人,并去收集那些已经写在各个频道和所有信息所在地的内容,而且我希望 AI 在汇总和收集数据方面多做一些重活。

I want you to start by collecting the agent by interviewing the people who hold the knowledge and harvest what's already written in all the channels and all the places where information already resides, and I want AI to do a lot of the heavy lifting in terms of aggregating and collecting the data.

所以对我们的客户智能体来说,我会确保销售、解决方案、客户成功和交付这些团队都贡献自己的一份力量。

So for our customer agents, for what I would do is I'll make sure that sales and solutions and success and delivery, they all will contribute their own piece.

然后我希望你进行优化。

And then I want you to refine.

我希望你整合来自不同来源的信息,把矛盾的地方摆出来。

I want you to merge the information coming from different sources, surface the contradictions.

我敢肯定你会发现真相有五个版本,然后给所有信息标上日期,把那些绝对不该分享的内容剔除出去。

I'm sure that you will find 5 versions of the truth and then date everything and keep out what should never be shared.

当然,这包括密码、关于他人的备注、一个客户的信息出现在另一个客户的空间里等等。

Of course, that includes passwords, notes about people, one customer's details in another customer's space, and so on.

然后我们需要审批它。

And then we have to approve it.

每一条信息都应该由它所属的人签字确认,然后 agent 的管理者把它们整合在一起。

Each piece should be signed off by whoever owns it, and the agent's owner puts it all together.

最后,这些信息很容易过期,所以你必须持续维护。

And lastly, this can go stale very quickly, so you have to maintain.

你需要决定 agent 可以根据自己的工作经验向记忆中添加哪些内容,以及哪些信息在进入知识和记忆之前必须先由人审核。

You need to decide what the agent may add to its own memory based on its working experience and what person has to review first when it goes into the knowledge and the memory.

我们希望确保,不会出现某个人对最新信息的定义悄无声息地变成所有人的标准,而大家并没有达成一致。

And we wanna make sure that one person's definition of what's latest quietly becomes everyone's without any agreement.

当然,还要把维护工作排入计划,因为你希望 agent 定期提出更新建议,并且需要有人来审核和批准这些更新。

And of course, put the upkeep on a schedule because you want the agent to propose updates regularly and a person needs to review and approve the updates.

这一点真的需要非常细心和自律,因为如果你在 agent 所依赖的知识的获取、维护和验证上没有一个足够好且能自我维持的流程,那它可能直接决定你的团队 agent 的成败,因为一旦出了问题,你没法快速修复。

And this is really the place to be very, very diligent and disciplined because it can truly make or break your team agent if you haven't done a good enough and self-sustaining process around acquiring and maintaining and verifying the knowledge that the agent taps into because it no longer serves you where you can very quickly fix anything that goes wrong.

如果你的团队 agent 对关键事项没有足够好的理解,它可能会给公司造成很大的混乱。

It can create a lot of havoc in your company if your team agent is not well educated enough on what matters.

第四个决定是 agent 能接触到什么,这是团队 agent 和个人 agent 的不同之处,因为个人 agent 代表你个人,而团队 agent 为很多人服务。

The 4th decision is what it can touch, and this is where team agents differ from the private ones because your own agent acts as you and a team agent acts for many people.

所以当然有大量的安全基础原则需要在这里应用。

So of course there are many security 101 that you need to apply here.

那些适用于所有 agent 的原则肯定要遵守,我重点说说针对团队 agent 的四条特定规则。

Those that apply to any agent, definitely stick for the 4 rules that are specific to, like, I'm gonna just stick to the rules that apply to team agents.

首先要决定的是它使用谁的权限,你有三种选择。

The first thing that you have to decide is whose access it uses, and you have 3 options.

你可以让它使用提问者的权限来操作。

You can use the access to act as whoever is asking.

这样它就只能看到提问者能看到的内容。

So it only sees what the person that was asking the question can see.

这可能是最安全的选择,但执行起来有时也是最复杂的,除非供应商已经帮你做好了。

And that's probably the safest choice, but sometimes the most complicated to execute unless the vendor already did it for you.

当团队成员的访问权限不同时,这可能是正确的方式。

And probably the right one when people on the team have different access levels.

另一个选择是为它设置一个独立的账号,只赋予工作所需的权限。

The other options that you have is to have its own account set up with exactly the access the job needs.

当整个团队都使用同样的共享资料时,这个方式就很合适。

And that's right when the whole team works on the same shared material.

最后,它也可以使用某个人的登录权限。

And lastly, it can use one person's login.

但只建议在只读和非敏感材料上这么做,因为所有与 agent 对话的人实际上都获得了那个人的权限。

Only ever do that for read-only and non-sensitive material because everyone who talks to the agent effectively gets that person's access.

所以我不建议走那条路。

So I would not recommend to go down that path.

对于我们的客户 agent,我可能会让它以提问者的身份操作,因为销售、客户成功和交付团队在 CRM 和其他系统中看到的内容各不相同。

And what I would probably do for our customer agent is to act as the person asking because sales, success, and delivery see different things in the CRM and in other systems.

所以我不希望 agent 向他们回复他们不应该看到的信息。

So I don't want to have the agents responding to them with information that they shouldn't be able to see.

这是关于 agent 能看到什么的第一条规则。

So that was the first rule on what the agent can see.

第二条规则是决定谁能向它提问。

The second rule is to decide who can ask it.

当 agent 有自己的账号时,所有能与它对话的人都可以使用这个账号。

And when the agent has its own account, everyone who can talk to it can use the account.

如果你把一个有权访问定价表的 agent 放在一个40人的频道里,其中还有几个承包商,那么突然之间,这40个人都可以通过询问来获取价格了。

And if you put an agent with access to the pricing sheet in a channel of 40 people, a few contractors among them, then all of a sudden all 40 can now get the pricing by asking.

这个 agent 知道的比一些能接触到它的人还多。

The agent knows more than some of the people who can reach it.

所以,决定谁能和它对话时,要像你决定它能看什么一样谨慎。

So decide who can talk to it with the same care you give to what it can see.

好吗?

Okay?

所以这是人们不注意时经常发生的事情。

So that's something that happens very regularly when people don't pay attention.

我还想让你决定答案落在哪里。

I also want you to decide where the answer lands.

这个正好相反。

This one runs the other way.

agent 使用提问者自己的权限,而提问者完全有权利问这个问题。

The agent uses the asker's own access and the asker has every right to ask the question.

问题是答案会出现在共享空间里,而那些没有权限的人也会看到。

The problem is that the answer shows up in a shared space in front of people who don't.

所以这是现在正在发生的事情。

So this is something that is happening right now.

如果你看 Claude 在 Slack 里的文档,它可以使用提问者在团队频道内的连接。

If you look at the documentation of Claude in Slack, it can use the asker's own connections inside the team channel.

在用户批准后,Anthropic 的文档目前指出,它不会考虑频道里还有谁。

And after the person approves, the Anthropic documentation currently notes that it doesn't consider who else is in the channel.

所以如果我批准使用我的连接器,获取所有我允许查看的信息,然后这些信息被扔进频道,其他人也能看到,这就是目前现有 Claude 实现的现实。

So if I approve the use of my connectors and fetch all the information that I am permitted to see, and then the information is thrown into the channel where others can see it, that's the reality currently with the existing Claude implementation.

所以如果信息敏感,答案必须私下发给提问者,而不是放在共享频道里。

So if it's sensitive, the answer has to go to the person who is asking privately and not in a shared channel.

最后,记录谁问了什么,这很重要,是日志记录。

And lastly, keep a record of who asked for what, and that's important logging.

因为当 agent 用自己的账户工作时,日志显示是 agent 做的,对吧?

Because when an agent works under its own account, the logs say the agent did it, right?

而你想知道是哪个人问的,这样你就能回溯,确保没有意外行为。

And you want to know which person asked so you can backtrack and make sure that there are no unexpected behaviors.

其他一切,比如从最小权限开始,以及让人批准任何不可撤销的操作,对所有 agent 都是一样的。

And everything else, like starting with least access and having a person approve anything that can't be undone, is the same for any agent.

所以我不是在给你做一对一的安全培训。

So I'm not giving you a security one-on-one.

好的。

Okay.

最后,最后一个决定,也是让你的团队 agent 活过第一周或第一个月的决定。

Lastly, the last decision and the one that will make your team agent live beyond its first week or the first month.

这就是完整的操作手册了,基本上。

That's the full, like, operating manual here.

当然,多人冲刺计划有更全面的思考方式,但这四点在实际中决定了 agent 的成败。

Of course, the multiplayer sprint has a much more comprehensive way of thinking about it, but these four points are what makes or breaks the agents in practice.

第一件事是只有一个负责人。

So the first thing is one owner.

团队中的任何人都能处理工作,但我希望有一个人或一个很小的团队来负责优先级、维护 agent,并在两个人提出相反要求时决定如何改进 agent 的知识或功能。

Anyone on the team can handle the work, but I want to have one person or a very small group of people who owns the priorities, maintains the agent, and decides, when two people ask for opposite things, how to evolve the agent knowledge or feature set.

他们还决定常设指令等等。

They also decide on standing instructions and so on.

这是第一点。

So that's one thing.

我想提的第二点是明确的互动规则。

The second thing that I want to mention is clear rules of engagement.

我希望你告诉人们如何使用它,它能做什么、不能做什么,它能看什么,谁能看到他们与它的对话,以及我们目前讨论的所有内容。

I want you to tell people how to work with it, what it does and doesn't do, and what it can see, who can see their conversation with it, and everything that we discussed so far.

我们还需要明确指示如何纠正 agent 的错误。

We also want to have clear indications of how to correct the agent when it's wrong.

而且我觉得,人们对他们正在使用的 Agent 和团队 Agent 的信任程度,会随着他们越了解 Agent 在学习什么以及学习过程而增加。

And I think that people trust the agents that they're using and the team agents the more they know what it's learning and what the learning process is.

接下来我想让你做的事情是,把决策放在它能看见的地方。

The next thing I want you to do is to put decisions where it can see them.

所以,一个团队 Agent 只能知道那些写在它能够访问的位置上的东西。

So a team agent only knows what's written down in a place that it can reach.

如果你的团队在私聊或者走廊里聊天中决定事情,Agent 永远不会知道这些。

And if your team decides things in private messages and hallway conversations, the agent will never hear about them.

要让 Agent 平稳运行,一部分工作就是让它能够实时了解团队中正在发生的事情,并确保团队决策和日常工作的进展能被 Agent 自己学到。

And part of running it smoothly is to have it be able to tap into what's happening in real time in the team and to make sure that the team decisions and the team's ongoing day-to-days are being learned by the agent itself.

最后,我希望你持续观察,因为你可能一开始会从很小的规模开始——理想情况下,你应该从一个小型试点小组开始,并准备几个测试问题,这样每当有变化时都可以重新跑一遍。

And lastly, I want you to keep watching because you will probably start with a small— ideally, you should start with a small pilot group and keep a few test questions that you can rerun whenever something changes.

但我也希望你持续监控,因为我们知道事情变化得非常频繁。

But I also want you to just monitor because we know that things change very frequently.

所以,不仅仅是当你想引入新模型、新工具、新知识或者指令上的新变化时要有合适的流程,还要监控一切是否正常运行。

So it's not just about having a proper process for whenever you want to introduce a new model or a new tool or new knowledge or new changes in instructions, but also just to monitor that everything is working properly.

当然,在某些情况下,如果团队 Agent 的行为不符合我们的预期,我们可能就需要把它淘汰掉。

And of course, in some cases, we would want to retire the team agent if it's not behaving properly as we expected.

我们讨论了很多内容,如果要在结束前做个总结的话,首先我希望我已经让你相信团队 Agent 对每个人都重要,即使你——可能需要一段时间才能真正构建一个,而且正确构建它们才是关键所在。

We covered a lot of ground, and if I need to pull it together before we close, first of all, I hope that I've convinced you that team agents matter for everyone, even if you're— it will take you a while until you actually are building one, and that building them properly is what makes the difference.

当然,并不是每个 Agent 都需要共享。

Of course, not every agent should be shared.

私人 Agent,或者共享知识和技能但配以私人 Agent,又或者是团队 Agent,这些都是可行的选择。

Private agent or shared knowledge and skills with private agents or team agents, these are all valid options.

并且在适当的地方使用它们。

And should be used where appropriate.

我们讨论了团队 Agent 的四种类型:专家、通用工作 Agent、桥梁和参谋长。知道你自己的使用场景属于哪一种,会告诉你很多关于如何把它做好的信息。

We talked about the team agents that are coming in 4 kinds, the experts, the common work agent, the bridge, and the chief of staff, and knowing which one your use case fits into tells you a lot about how to get it right.

我们还讨论了什么时候该等待的三个信号:是品味胜过标准,还是没人能拥有知识,或者是它并没有简化工作。

And we also talked about 3 signs on when to wait, whether it's because taste beats the standards, whether because nobody can own the knowledge, or when it doesn't simplify the work.

而当你开始构建时,我们按照顺序讨论了五个决策:它做什么、它住在哪里、它知道什么、它能触碰什么,以及你如何运行它。

And once you're building, we went over the 5 decisions in order of what it does, where it lives, what it knows, what it can touch and how you run it.

如果你记住这些,你就有了正确构建团队 Agent 所需要的东西。

And if you take those with you, you have what you need in order to get the team agent properly.

如果你想具体操作,有一个快速的方法。

If you want to concretely do that, a quick way to do that.
F1
F137:09

最后,一切都在变化。

And lastly, everything is changing.

很可能每周我们都会有相关的发布,等你们听到这期节目时,可能已经有东西发布了。但我确实认为今天讨论的所有内容,无论接下来推出什么,都依然成立,因为当一个 Agent 被正确构建、决策得当,它就和任何特定的工具或功能是正交的。

Odds are that every week we will get a relevant release, and by the time you hear this, maybe something has already been released, but I do think that everything that we cover today holds no matter what ships next, because when an agent is built properly and the decisions are made right, it's orthogonal to any specific tool or feature.

业务决策和团队标准化比工具重要得多,而随着更多发布,工具只会让事情设计得更好。

It's the business decision and the team standardization that matters much more than the tools that will help make it better by design, the more releases we will have.

如果我要对今年余下的时间做些预测,我认为我们会看到今天所讨论的内容越来越正式化,也会有更多工具和功能帮助我们在身份、权限、所有权等方面做得更好更简单。

And if I need to make some predictions for the rest of the year, so I think that we will see more and more formalization of what we just covered and more tools and features that will help us get it even better and easier around identity, permissions, ownership, and so on.

同时,我们总能信赖实践者们在公开场合分享他们的很多经验教训,这样我们就能从许多其他 AI 前沿领域的个人和公司那里学到东西,看看他们是怎么做的。

As well as like, we can always trust the practitioners to share many of their learnings in the public eye so we can learn from many of the other AI frontier individuals and companies and see how it's working for them.

就这些了。

That's it.
M1
M138:11

太棒了。

Awesome.

内容很棒,Noufar。

Great stuff, Noufar.

在结束之前,我有几个想法想以讨论的方式抛出来。

I have a few things that I want to lob out there discussion style just as we close out.

首先,我想问的是,你给出了四种你见过的 Agent 原型。

First of all, I guess the question, you gave 4 archetypes of different types of agents that you've seen.

根据你的观察,团队起步时,有没有哪种原型比其他的更常见?

Do you see, are any of them more common starting places than others for teams that you've observed?
F1
F138:29

我觉得从理论上讲,那种存在于个人之间或团队之间的 Agent 定义是最吸引人的,但也是执行起来最困难的。

I think that in theory, your definition of an agent that lives between individuals or between teams sounds the most attractive, but it's the hardest to execute.

所以我觉得这些其实是我看到的,不太会优先去尝试的。

So I think those are actually the ones that, from what I'm seeing, are not the first to go for.

我见过很多非常成功的第一个版本,就是那种专家型智能体,帮团队创造更多冗余,我说的冗余是好的那种,还能减轻公司里那些瓶颈环节的压力。

And I've seen various very successful versions of the first one, of the expert agents that help create more redundancy in a team, redundancy in the good sense, and relieve some of the burden on those bottlenecks within the company.

这类我已经看到很多实际落地了,而且我觉得工具越容易上手,这些就越容易搭建。

So those I've seen a ton of implementations already, and I think the more the tools make it more accessible, the easier those will be to build.

所以这些大概是最容易摘的果子。

So those are probably the lowest hanging fruits.
M1
M139:08

有意思,因为那正好是我脑子里想的。

That's funny because that's exactly where my head goes.

我个人特别被那些我觉得最难做的吸引。

I'm super attracted to the ones that I think are most difficult to build.

我也见过很多那种专家型的简单实现——其实我们聊了很久都没意识到它算这种团队智能体——就是代理化的团队内部知识库,比如各种政策,像早退啊之类的,就是公司数据库里的信息,你可以通过聊天机器人直接查。

I've also seen a lot of, you know, a very simple implementation of that expert one, which, you know, we've talked about for a long time without even identifying it as this sort of team agent is just basically the agentified team internal knowledge hub, you know, policies around whatever, like early dismissal, who knows, like whatever, you know, that just the company database of information that you can access through the chatbot instead.

很多公司从一开始就在这个用例上取得了巨大成功,它简单、快速、容易上手。

That's been something that companies have had a ton of success with as just an early, easy, fast use case right from the beginning.

我还有个问题想问你,就是肯定有很多人在犹豫,在 GrokBot、Microsoft Copilot 或者他们用的那些核心工具——比如 OpenAI、Anthropic——还没推出原生版本之前,他们该投入多少来搭建这些东西?

Another question that I have for you is how, you know, I'm sure that a lot of folks are sitting there wondering how much they should invest in building these sorts of things before GrokBot or Microsoft Copilot or one of these sort of core tools that they might be using, OpenAI, Anthropic, just drop the sort of native version of this.

而且很明显,有些迹象表明他们正在往这个方向想。

And there's clearly some indications that they're thinking in this way.

我觉得 Claude Tag 是目前最好的例子。但问题是,现阶段深入做这件事的价值是不是在于让你理解理论、学会定制化,等未来更好的工具出来时就能用上?还是你怎么看这个取舍?

I think Claude Tag being the best example so far, but, you know, is this one where the value of digging in at this stage is going to be so you understand the theory and the ways to customize when better tools come around in the future, or how do you think about that trade-off?
F1
F140:29

我觉得最核心的工作永远是配置和知识梳理。

I think the heavy lifting is always gonna be the configuration and the knowledge curation.

所以我会选一个跟你现有工具生态最接近的工具,然后想办法把其他功能都实现出来。

So I would select the one tool that is adjacent the most to your existing tool ecosystem and figure out how to implement all the rest.

如果有更好更新的工具出现,你也能准备好,因为你们已经在基础事实、这些智能体的禁忌、使用场景上达成共识了。

And if like new, better improved tools come to play, you'll be ready because you'll already agree on the ground truth, on the dos and don'ts of these agents, on the use cases.

所以哪怕一开始实现得很笨拙,也能为未来打好基础。

So even if you at first implement them very naively, that's gonna have your future ready.

至于要不要自己从头搭建一个类似竞争产品或者团队级别的框架,如果你有那个本事,做起来不难,也有合理的理由,那你可以做。

As for going and building your own, like a competing product or your own like team-level harnesses and so on, if you have the chops and you can do that easily and you have the justification, you can.

但我现在不会把精力花在自建一个 Claude Tag 或类似的东西上,因为我们俩都同意这些功能都会来,而且很可能变得非常易用、非常智能。

But I'm not sure that I would've spent my energy now on going and building my own, like my own version of ClaudeTag or similar when we were, I think both of us agree that all of these are coming and will probably be made very accessible and very smart.

你的护城河很可能在于所有那些这些公司无法掌控的东西上。

And your moat is probably in everything that these companies cannot own.
M1
M141:31

嗯,老样子,感谢又一场精彩的 Operator's Cut,期待你很快再回来。

Well, thanks as always for another great Operator's Cut, and excited to have you back soon.
F1
F141:35

谢谢。

Thank you.
M1
M141:35

拜。

Bye.
已剔除 4 处广告(点击展开查看)
M1
M11:12广告 · 已剔除

First of all, thank you to today's sponsors, KPMG, Blitzy, Harbor, and HyperAgent.

To get an ad-free version of the show, go to patreon.com/AIDailyBrief, or you can subscribe on Apple Podcasts.

And to learn more about sponsoring the show, send us a note at [email protected].

M1
M11:40广告 · 已剔除

First of all, this Thursday, October 1st, we have a free live webinar all about building your personal AI benchmark.

The whole idea is that when you get a new model like Opus 5.5 or Sonnet 5.5 or Gemini 4 or whatever model comes next, this will help you put together your own standard benchmark to better understand where that model is going to fit into your own process.

That is completely free, and if you register, you will get all the materials after, even if you can't attend.

Again, that is coming up this Thursday, October 1st.

Now, speaking of training, if you want to go a little bit deeper, the next cohort of our Superintelligent Executive AI and Agent training programs is coming up.

The Executive Agent Leadership program is where you learn how to build AI agents for real business needs.

As well as building a playbook to scale them safely across your organization.

And if you feel you need a little bit more background before you get into that, you can also do the Executive Catch-Up Program.

The next Agent Leadership Cohort starts on October 5th, while the next Executive Catch-Up Program starts a week later on October 12th.

M1
M19:43广告 · 已剔除

A new study from KPMG and the University of Texas at Austin found that when people work with AI, similar skills don't guarantee similar outcomes.

Researchers studied more than 500 early-career professionals and found that the best performers consistently amplified the value of AI by guiding, evaluating, and refining its outputs.

These top performers, called AI amplifiers, weren't defined by what they knew alone, but by how they worked with AI.

Learn more about what separates AI amplifiers from everyone else at kpmg.com/us/aiamplifiers.

Blitzy deeply understands your codebase before it writes code.

Here's the first place that pays off: security in the age of AI.

Vulnerabilities don't live in isolation— they live buried inside millions of lines of interconnected code, where patching one thing quietly breaks 3 others.

That's why surface-level scans fail.

Blitzy starts from its knowledge graph of your entire application, identifies and surfaces CVEs across the full estate, proactively recommends patches, and can execute the PR.

Each fix is grounded in how your systems connect and validate so nothing new breaks, and the knowledge graph dynamically updates, keeping you ahead of an ever-accelerating threat landscape.

One Blitzy customer resolved 21 active CVEs across 6 core microservices in 4 days.

Zero compile errors, every validation scan clean, months of planned work fixed in less than a week.

Security remediation grounded in real architectural context at the speed of compute.

Harden your codebase at blitzy.com.

That's blitzy.com.

Every episode, I talk about the competition between OpenAI, Anthropic, xAI, Google, and Meta.

And if you've been listening for a while, you might have a favorite.

Maybe you think OpenAI and Anthropic can stay ahead, or perhaps Meta's open-source strategy can win out.

Whatever your view, every AI lab creates a different investment opportunity.

Harbor Capital Advisors' AI Lab Ecosystem ETF suite lets you invest in the ecosystem behind the AI lab you believe in.

Search Harbor AI Lab Ecosystem ETFs wherever you invest or follow @HarborCapital on X to learn more.

Visit harborcapital.com for a prospectus containing investment objectives, risks, fees, expenses, and other important information.

Read and consider it carefully before investing.

Includes principal loss and artificial intelligence-related risks.

Harbor ETFs are distributed by Foresight Fund Services, LLC.

Harbor is not affiliated with AI Daily Brief, and the funds are not affiliated with, sponsored by, or endorsed by any AI lab.

This is a paid advertisement and not personalized investment advice.

Investing involves risk, including possible loss of principal.

This episode of the AI Daily Brief is brought to you by HyperAgent, where you run fleets of agents your team can manage together.

Forget local agents and chat workflows waiting on your laptop to be prompted.

HyperAgent deploys always-on agents in the cloud, doing real work across the tools your team already uses.

Marketing agents turn competitor moves into landing pages.

Sales agents enrich leads, draft emails, and update the CRM.

Ops agents chase the paperwork and track the budget.

Every agent has access to shared context and follows your rules about scope and approvals.

It's time you had agents that feel like teammates.

Hire yours at HyperAgent.

Get $100 in credits at hyperagent.com/AIDailyBrief.

F1
F136:27广告 · 已剔除

So first of all, we have the Multiplayer AI Sprint, which is free and will walk you through a team activity over 4 weeks configuring everything that we discussed, some of them in greater detail.

If you want to learn how to properly build serious team agents and agent rosters and how to do that in the best possible way, we have another cohort of the Executive Agent Leadership that starts on October 5th, and we'll be happy to see you with all of our builders.

However, if you feel that you need a little bit of a catch-up before you go and build agents for teams and rosters of agents and so on, we also have the Executive Catch-Up that helps you become best-in-class AI user before you go and build those agents.