Practical AI
From AGENTS.md to Enterprise Deployment
从 AGENTS.md 到企业级部署
00:002931
00:00
Welcome to the Practical AI Podcast, where we break down the real-world applications of
欢迎来到 Practical AI Podcast，我们会拆解 AI 的实际应用，以及它如何塑造我们生活、工作和创作的方式。
00:07
artificial intelligence and how it's shaping the way we live, work, and create.
我们的目标是帮助让 AI 技术对每个人都实用、高效、触手可及。
00:13
Our goal is to help make AI technology, practical, productive, and accessible to everyone.
无论你是开发者、企业领导者，还是只是对热潮背后的技术感到好奇，你来对地方了。
00:18
Whether you're a developer, business leader, or just curious about the tech behind the buzz,
一定要在 LinkedIn、X 或 Blue Sky 上联系我们，及时了解新节目上线、幕后内容和 AI 洞察。
00:23
you're in the right place.
你可以在 practicalai.fm 了解更多。
00:25
Be sure to connect with us on LinkedIn, X, or Blue Sky to stay up to date with episode
一定要在 LinkedIn、X 或 Blue Sky 上关注我们，及时了解新一期
00:29
drops, behind the scenes content, and AI insights.
上线、幕后内容以及 AI 洞察。
00:32
You can learn more at practicalai.fm.
你可以在 practicalai.fm 了解更多。
00:35
Now, onto the show.
现在，进入正题。
00:41
Welcome to another episode of the Practical AI Podcast.
欢迎来到另一期Practical AI Podcast。
00:45
This is Daniel Whitenack.
我是Daniel Whitenack。
00:46
I am CEO at Prediction Guard, and I'm joined as always by my co-host, Chris Benson,
我是Prediction Guard的CEO，和往常一样，我的搭档主持Chris Benson也在这里，
00:52
who is a principal AI and autonomy research engineer.
他是一位首席AI与自主研究工程师。
00:56
How you doing, Chris?
你好吗，Chris？
00:56
Hey, doing great today.
嘿，今天挺好的。
00:57
How's it going?
你怎么样？
00:58
It's going great, and I know this is going to be a great conversation for a few different
一切都很顺利，而且我知道，出于几个不同的原因，今天会是一场很棒的对话。
01:04
reasons today.
其中一个原因是，我们的嘉宾是即将举办的 Midwest AI Summit 的演讲嘉宾之一，那场活动会非常精彩。
01:05
One of those being our guest is one of the speakers at the upcoming Midwest AI Summit,
到时候会有一大堆很棒的演讲嘉宾。
01:11
which is going to be amazing.
我会去做一点 MC，看看自己会不会搞砸。
01:13
There's going to be a bunch of amazing speakers there.
我想鼓励我们的听众去看看。
01:15
I'm going to do a bit of MC-ing, see if I don't mess that up.
10 月 15 日在 Indianapolis，用 Practical AI20 可以打八折，所以去看看吧。
01:19
I would encourage our listeners to check that out.
我也想鼓励我们的听众去看看。
01:22
October 15th in Indianapolis, you can get 20% off with Practical AI20, so check that out.
10 月 15 日在 Indianapolis，用 Practical AI20 可以打八折，所以去看看。
01:30
Amazing speaker, going to be at the Midwest AI Summit.
很棒的演讲嘉宾，会参加 Midwest AI Summit。
01:33
Also a fellow podcaster, which is great.
也是同行播客主，这很棒。
01:37
One of the hosts of Cloud Foundry Weekly and a tech marketing whiz at Broadcom, Tanzu,
Cloud Foundry Weekly 的主持人之一，也是 Broadcom、Tanzu、
01:47
VMware.
VMware 的技术营销高手。
01:48
Welcome, Nick.
欢迎，Nick。
01:49
Great to have you.
很高兴你能来。
01:50
Thanks for having me.
谢谢邀请我。
01:51
Great to be on the show.
很高兴能上这个节目。
01:52
Yeah.
是啊。
01:53
Excited.
很期待。
01:54
Also excited to see all in person in Indianapolis.
也很期待在 Indianapolis 线下见到大家。
01:57
Yeah, it's going to be fun.
是啊，肯定会很好玩。
01:58
We're representing the good representation of the Silicon Prairie here.
我们在这儿代表着 Silicon Prairie 好的一面。
02:03
I like it.
我喜欢。
02:05
So yeah, Nick, I know you're going to be talking, one of the things I thought was cool about
所以，对了，Nick，我知道你要演讲，我觉得很酷的一点是
02:12
your talk at the Midwest AI Summit is you bring in this concept of taking agents.md and
你在 Midwest AI Summit 的演讲里引入了这个概念：把 agents.md 拿来，然后
02:19
shipping that to a production environment where you run an actual agent in an, quote,
把它发布到 production environment，在那里你会在一个，quote，
02:26
enterprise environment, which seems to be really fascinating.
enterprise environment 里运行真正的 agent，这看起来真的很有意思。
02:31
We recently had on someone from the agent AI foundation who is kind of stewarding the
我们最近请了一位来自 agent AI foundation 的人，他有点像是在负责管理
02:40
model context protocol and agents.md and those sorts of things.
model context protocol 和 agents.md 以及诸如此类的东西。
02:44
So very relevant kind of carry on from that conversation.
所以非常相关，算是接着那场对话往下聊。
02:48
But I'm wondering as we set that up, what exactly does it mean to be operating in an enterprise
但我在想，当我们把这个搭起来的时候，在 enterprise
02:57
environment?
environment 里运营到底意味着什么？
02:58
So the customers that you're working with that you're kind of day to day, what makes an
所以你现在日常打交道的那些客户，是什么让一个
03:03
enterprise environment an enterprise environment?
enterprise environment，是一个 enterprise environment 吗？
03:06
What are some of the concerns or characteristics of that type of environment?
这类环境会有哪些顾虑或特点？
03:11
And I know you're always experimenting.
而且我知道你一直在做实验。
03:15
You do a lot of home lab stuff as well.
你也做很多 home lab 的东西。
03:18
So always have a whole range of experience in working in different sorts of environments.
所以你总是在各种不同类型的环境里工作，经验非常丰富。
03:24
So yeah, if you could just help us understand from a general perspective, what does that
所以，嗯，你能不能从整体角度帮我们理解一下，那
03:27
mean when you say kind of enterprise?
当你说所谓的 enterprise 时，是什么意思？
03:30
Right.
对。
03:31
So yeah, if we think of the enterprise, we think about a few things.
所以呢，说到企业，我们会想到几件事。
03:36
One, let's just assume you're going to have not limited or no internet access.
第一，我们就假设你要么有不受限的 internet access，要么根本没有。
03:42
And I've been at VMware Tanes here for about five years.
而且我在 VMware Tanes 这里大概待了五年。
03:47
And before that, I was at one of these large enterprises for about 14 years.
在那之前，我在这些大型企业中的一家待了大概 14 年。
03:51
So when we had vendors come in, they would always assume that we could just go to the internet
所以当供应商来的时候，他们总是假设我们可以直接访问 internet
03:56
and it was always like, well, try again, because that's not going to work here.
而我们总是会说，好吧，再试试吧，因为在这里行不通。
03:59
So that's kind of always like the biggest hurdle where you're going to be regulated and
所以这差不多一直是个最大的障碍，因为你会受到监管，而且
04:05
or you're not going to, what you can do at home is not going to be what you can do at work,
或者你做不到，你在家能做的事，不会等于你在工作中能做的事，
04:12
so to speak.
可以这么说吧。
04:13
So you're going to be highly regulated.
所以你会处在高度监管之下。
04:14
You're going to have to deal with potential like PCI or SOX compliance or HIPAA or FIPs and
你得应对像 PCI 或 SOX compliance、HIPAA 或 FIPs 这些潜在要求，而且
04:22
I could probably go on and on about all the different compliance tiers that you would
我大概能没完没了地说下去，说你将
04:25
have to deal with.
不得不应对的各种不同 compliance 层级。
04:26
So it's just a whole different ballgame where you could potentially do something that has
所以这完全就是另一码事，你可能会做出一些会带来
04:30
dramatic consequences.
巨大后果的事情。
04:33
And you have a highly controlled, highly regulated environment where even some of the customers
而且你身处一个高度受控、高度监管的环境，甚至有些客户
04:37
on work with, they actually have the real air gap where we're carrying in things, physically
跟他们合作的时候，他们其实有真正的 air gap，我们得把东西物理地
04:43
into the data center because there is no internet access type of thing.
带进 data center，因为那里没有 internet access 这种东西。
04:47
So it's a wild world and it frustrates a lot of, I think, new folks to the enterprise
所以这是个挺野的世界，我觉得很多刚接触 enterprise
04:54
space.
space 的新人都会挺受挫的。
04:55
But if you've been in in a while, you know how to deal with it and how to handle it type
但如果你已经在这一行待了一阵子，你就知道怎么应对、怎么处理
04:58
of thing.
这类事情。
05:00
And I guess, you know, people have been dealing with these types of environments for quite
而且我猜，你知道，人们应对这类环境已经
05:06
a while, right?
挺久了吧？
05:07
Because there have been enterprises for quite a while.
因为企业已经存在相当长一段时间了。
05:11
Maybe just give us a sense of like leading up to AI agents and we'll talk about agents
也许先给我们讲讲，比如说在 AI agents 出现之前，我们接下来会聊到 agents
05:18
here in a second, but leading up to AI agents like what were some of the ways that you
马上就会聊，但在 AI agents 之前，你们当时有哪些方式……
05:22
could, I guess, ease, you know, put some ointment on those, those pains of working in that
我觉得，这也许能缓解一下，你知道，给在那种环境里工作的那些，那些痛点抹点药膏。
05:29
sort of environment.
你会怎么把一个 application 放进那种环境里，或者让它变得 enterprise ready？
05:30
How would you get an application into that sort of environment or have it be enterprise
所谓的。
05:36
ready?
所以开发者就不用担心，比如，你知道，给 container 做安全防护，或者类似的事情。
05:37
Quote unquote.
platform 会负责这些。
05:38
And then we can shift and talk about, you know, what changes with agents.
然后我们可以转过来谈谈，你知道，agents 会带来什么变化。
05:42
Yeah.
对。
05:43
So, I mean, I think, so, you know, I currently work for VMware TANZU and we run a platform
所以，我是说，我觉得，所以，你知道，我现在在 VMware TANZU 工作，我们运营一个 platform
05:48
or platform as a service, you know, geared for private cloud, whether that be on your,
或者说 platform as a service，你知道，是针对 private cloud 的，不管是在你的，
05:54
like on your own bare metal in your data center, or you could deploy a private cloud in your
比如在你自己的 data center 里的 bare metal 上，或者你可以在你的
05:58
own, like VPC on a hyperscaler, that type of thing.
自己的，比如 hyperscaler 上的 VPC 里部署 private cloud，那种事情。
06:03
But we think about that, you know, we want to, in the scope of an enterprise, right?
但我们考虑的是，你知道，我们想在 enterprise 的范围内，对吧？
06:07
You have all these regulations, controls, et cetera.
你有所有这些法规、管控等等。
06:09
You want to make that path to production, kind of the least resistant and easiest path.
你想让那条通往 production 的路径，变成阻力最小、最好走的那条路。
06:15
And if you make that path to production, like all the checkboxes and everything good
而如果你把那条通往 production 的路径，所有该勾的 checkboxes 和一切都弄到
06:18
to go, that's going to be a highly adopted path.
准备就绪，那就会是一条会被高度采用的路径。
06:21
So the whole concept of, you know, TANZU platform, and it's actually based on an open source
所以，你知道，TANZU platform 的整个概念，实际上它是基于一个 open source
06:27
project called Cloud Foundry, which is, I think, I mean, it's, you know, predates Kubernetes
项目，叫做 Cloud Foundry，我觉得，我是说，它，你知道，比 Kubernetes
06:31
and Docker.
和 Docker 还要早。
06:32
So, you know, 2011, it's just when it kind of started originally out of VMware and spun
所以，你知道，2011 年，就是它最初从 VMware 里开始，然后分拆
06:35
out and then became part of an open source project.
出来，后来成为某个 open source 项目的一部分。
06:38
So it's been around and battle tested and seen a lot of things where you have that concept
所以它已经存在了一段时间，经过了实战检验，见识了很多东西，就是你有那种概念，比如我就想把我的代码发到平台上，平台就会知道，比如最佳实践是什么，它会为开发者构建一个container。
06:43
of, like, I just want to take my code and send it to the platform and the platform is
所以开发者就不用担心，比如，你知道，给container做安全加固之类的。
06:46
going to know, like, the best practices for, it's going to build a container for the
平台会搞定这些。
06:50
developer.
它会处理，你知道，ingress，比如search certificates，所有，你知道，load balancers，
06:51
So the developer doesn't have to worry about, like, you know, securing the container
所以开发者就不用操心，比如，你知道，给 container 做安全防护之类的事情。
06:54
or anything like that.
平台会搞定这些。
06:55
The platform would take care of that.
平台会处理那个的。
06:57
It will handle, you know, ingress, like, search certificates, all, you know, load balancers,
它会处理，你知道，ingress，比如 search certificates，还有所有那些，你知道，load balancers，
07:03
all those things just handled health monitoring, potentially even kind of like sandboxing the
所有这些事情只是处理了 health monitoring，甚至可能有点像在做 sandboxing，把
07:07
apps from each other, something that, you know, maybe some of these labs may learn
apps 彼此隔离开来，这件事，你知道，也许其中一些实验室可以学习
07:12
from, but so that the apps can't escape type of thing.
从中，但这样 apps 就不能逃逸之类的。
07:18
And you know, if the app needs services like a database or messaging, middleware, any
而且你知道，如果 app 需要像 database 或 messaging、middleware 这样的服务，任何
07:23
type of thing, even maybe a large language model, it can, what we call bind or it kind
类型的东西，甚至可能是一个 large language model，它就可以，我们称之为 bind，或者它有点
07:28
of connects to a service on the fly and that that's all kind of handled under, underneath
动态连接到一个服务，而这一切都在某种程度上被处理在，在
07:32
the cover.
幕后。
07:33
So it's just a few commands like you push, bind the scale, your app with just a few simple
所以只需要几个命令，比如你 push、bind、scale 你的 app，只需几个简单的
07:39
constructs.
constructs。
07:40
And that's, that's like, that's how you go to production.
这就是，这就像，这就是你走向 production 的方式。
07:42
And what we see with enterprises, once that path, like, once they take this, like,
而我们看到在企业里，一旦这条路，就像，一旦他们走上这条，就像，
07:46
this is a tainted platform, it's been certified, they're all these different standards.
这是一个经过认证的平台，它已经获得认证，有各种不同的标准。
07:50
It's just kind of like an unlock for these enterprises because it's so easy and so simple
这有点像是对这些企业的一种解锁，因为它用起来太容易、太简单了，
07:56
to use and they don't have to, you know, they don't have to recertify everything.
他们不用，你知道，他们不用把所有东西重新认证一遍。
07:59
It's just like this framework that you can use to get apps into production.
它就像这个框架，你可以用它把 apps 推向 production。
08:03
And then the biggest thing is that you want these, the enterprises or the businesses,
然后最重要的是，你希望这些，这些企业或者这些公司，
08:07
they don't want their developers, you know, hand crafting a new platform or a new way to
他们不想让他们的开发者，你知道，手工打造一个新的 platform，或者一种新的
08:12
deploy apps for every sub team and have like 100 different snowflakes, right?
给每个子团队 deploy apps 的方式，然后搞出像 100 个不同的 snowflakes，对吧？
08:16
They want the same repeatable pattern across the board so that they can audit it and secure
他们想要所有地方都用同一种可重复的 pattern，这样他们就能 audit 它、secure
08:20
it and have their developers actually spend time writing the business logic versus like
它，并且让他们的开发者真正花时间写 business logic，而不是像
08:25
fiddling with infrastructure.
瞎折腾 infrastructure。
08:26
So that's kind of like the whole premise behind, you know, the whole platform is sort
所以这差不多就是，你知道，整个 platform 背后的整个前提，就是这个 platform 有点
08:29
of just construct and, you know, tainted platform.
像是 just construct，而且，你知道，tainted platform。
08:32
And as you can imagine, or go ahead and yeah, no, no, go ahead, finish up and then I'll
而且你可以想象，或者你继续，对，不，不，你继续，说完然后我就
08:36
follow all good.
都照做就好。
08:37
Okay.
好。
08:38
As, as you could imagine, those principles when we're talking about apps could probably
就像你能想象的那样，我们在聊 apps 的时候，那些原则很可能
08:42
play nicely into this agent world that we're living in.
能很好地融入我们正处的这个 agent 世界。
08:45
So it sounds like you're drawing a lot of commonality between kind of traditional app development
所以听起来，你是在把很多共通点，放在某种传统的 app 开发
08:51
and all of the structure that we're all used to, that we've been doing forever.
和我们早已习惯、一直都在做的那整套结构之间。
08:59
And kind of now we're into this agentic world.
而且现在，我们算是进入了这个 agentic 世界。
09:02
And maybe people have been thinking about that in a different way, but it sounds like
也许人们一直是以不同的方式在想这件事，但听起来
09:06
you've really kind of said and correct me if I'm wrong, I'm going to get to throw something
你其实某种程度上已经说到了，如果我理解错了请纠正我，我想在这里抛个想法出来。
09:10
out here.
某种意义上这有点像同一回事，虽然有区别，但 agents 和 apps 大概应该按同样的方式对待。
09:11
It's kind of the same thing in the sense of there are differences, but agents and apps
我理解得对吗？就你构建这件事的方式来说，是不是说，也许人们需要回头看看自己做了很多年的东西，然后让那些也能以同样的方式适用于 agents？
09:17
should sort of be treated the same way.
是这样吗，这么说合理吗？
09:19
Is that, am I getting that correctly in terms of how you're structuring that and that
是这样吗，我理解得对吗，就你是如何构建那个以及那个来说
09:24
maybe people need to look at what they've been doing for years and make that work for agents
也许人们需要看看他们多年来一直在做的事，并让那套东西能为 agents 所用
09:30
in the same way?
以同样的方式？
09:31
Is that, is that fair?
这，这公平吗？
09:33
Yeah.
嗯。
09:34
I mean, I think it's a pretty fair assessment.
我是说，我觉得这个评价挺公道的。
09:35
And we're starting to see that more and more just within our customer base.
而且在我们自己的客户群里，我们已经开始越来越常看到这一点了。
09:39
And even if you think about apps back in the day, like someone would write some app on
而且就算你想想以前的 apps，比如有人会写个 app 在
09:45
their laptop and run it in their laptop, but it's like, we can't just run it out.
自己的 laptop 上，然后就在自己的 laptop 上跑，但这就好像，我们不能只是让它跑在本地。
09:48
You've got to get it off your laptop somewhere and the same thing with agents, you're like,
你得把它从你的 laptop 上弄到别的地方去，agents 也是一样，你会想，
09:51
I got this.
我能搞定。
09:52
If I got all these agents from my laptop, if I shut my lid and everything stops, I don't
如果我把所有这些 agents 都跑在我的 laptop 上，要是我一合上盖子，一切就都停了，我不……
09:56
really want that to happen.
真的希望那能发生。
09:57
I want the run 24-7 365 and do all the work I give them.
我希望它们 24-7 365 地运行，并完成我交给它们的所有工作。
10:01
So yeah, I think that's kind of precisely what we're going with.
所以对，我觉得这差不多正是我们打算采用的方向。
10:05
So I guess then it begs the question, why are people kind of throwing out what maybe
那我想，这就引出了一个问题：为什么人们有点在抛弃那些也许
10:11
they're intuition?
是他们的直觉的东西？
10:14
In certain ways, people are saying, oh, we have agents now.
在某些方面，人们在说，哦，我们现在有 agents 了。
10:17
We need to do something totally different, which is maybe driven by some things that make
我们需要做一些完全不同的东西，而这也许是由一些让
10:24
an agent not just a sort of simple application, but also it doesn't necessarily mean, obviously
agent 不只是一个简单应用的因素所驱动的，但显然，它也不一定意味着
10:33
we have to throw out everything we've learned about platform engineering and how to deploy
我们得把过去学过的所有 platform engineering 以及怎么 deploy 东西的那一套全都扔掉。
10:38
things.
所以从你的思路来看，agent 就等于另一个 app 这种想法，在什么地方成立？
10:39
So from your mindset, where does the kind of agent equals just another app?
它又可能在哪儿行不通，或者说，在 agent 这一侧、在 deployment pathway 上，你要面对的那些独特特点是什么——是你以前可能根本不用考虑，或者至少现在要用新方式去考虑的？
10:47
Where might that fall apart or what are the unique characteristics that you're dealing
嗯，我觉得一些最大的区别在于，当你谈到 Cloud Foundry 和 TNC platformer 时，它某种程度上是基于 12-factor application 这个前提。
10:51
with on the agent side, on the deployment pathway that you might not have had to consider
而在 agent 这边，在 deployment pathway 上，那些你以前可能不需要考虑的
10:58
before or at least are considering in a new way?
以前不用考虑，或者至少现在是以一种新的方式在考虑？
11:00
Well, I think some of the biggest differences where you talk about Cloud Foundry and TNC platformer
嗯，我觉得一些最大的差异在于，当你谈到 Cloud Foundry 和 TNC platformer 的时候
11:06
was kind of based on this premise of a 12-factor application.
这算是基于 12-factor application 这个前提。
11:09
We're like storage and state are a little decoupled cleanly from the actual apps.
我们有点像，storage 和 state 跟实际的 apps 算是比较干净地 decoupled 了。
11:15
You get like, you know, I can scale up to a thousand instances and be fine in your
你会发现，你知道，我可以 scale up 到一千个 instances，然后在你
11:19
handle sessions, dating things.
处理 sessions、dating 这些东西上也没问题。
11:20
The traditional harnesses or agents were kind of built originally just for like, ah,
传统的 harnesses 或者 agents 最初被搭建出来，差不多就是为了那种，啊，
11:24
I've got file system access and I can write about your MD files and that's like my memory.
我有 file system access，可以写关于你的 MD files，那就像我的 memory。
11:28
Everything's great.
一切都很好。
11:29
So you start to get into a Cloud world that things spin up and down in our ephemeral.
所以你就开始进入一个 Cloud 世界，东西在里面 spin up 又 spin down，而且是在我们 ephemeral 的环境里。
11:34
You're going to want to save those MD files somewhere they're going to go off into
你会想要把那些 MD files 存到某个地方，它们会跑进……
11:40
the ether.
the ether。
11:41
So that's probably the, I think the biggest challenge when we start talking about how do
所以这大概就是，我觉得，最大的挑战，就是当我们开始讨论我们怎么
11:44
we take the app deployment methodology, right, and then push that into like an agent running
把 app deployment methodology 拿过来，对吧，然后把它推进到类似一个 agent 运行在
11:50
on a kind of a Cloud platform per se.
某种 Cloud platform 上，就它本身而言。
11:53
It's probably the biggest challenge.
这大概就是最大的挑战。
11:54
And then just kind of getting the fact that there's some valid use cases for like your laptop
然后还得认识到这个事实，就是有些 valid use cases 是给比如你的 laptop 用的
11:59
or whatever.
或者别的什么。
12:00
But then also as we're starting to see more and more, like enterprise was, enterprises want
但然后，随着我们开始越来越多地看到，比如 enterprise 以前是，enterprises 想要
12:04
agents to be on demand and like usable kind of in a, let's say in a CICD pipeline or like
agents 要能 on demand，而且有点像能用在，比如说，一个 CICD pipeline 里，或者那种
12:09
on, you know, within part of their, you know, ecom or suite or just normal applications and
然后，你知道，在他们那部分里，你知道，ecom 或 suite，或者就是普通应用里，然后
12:14
you want those to be kind of really close to the apps, right, because if you're, if you
你会想让这些离 apps 特别近，对吧，因为如果你，如果你
12:18
have them way off, you know, even just from a pure networking perspective, if you have,
把它们放得特别远，你知道，哪怕就从纯 networking 的角度看，如果你有，
12:22
you know, the agent or the LLM or whatever, that's like, you know, 30 hops away from the
你知道，那个 agent 或者 LLM 或者随便什么，它就像，你知道，离那
12:26
microservices, try to call it, there's, you know, physics that are going to add on to
microservices 有 30 hops 那么远，试着调用它，你知道，会有物理规律加在
12:30
all of that latency and potentially at scale, even be problematic.
所有这些 latency 上，而且可能 at scale 的时候，甚至会成问题。
12:33
So they, they kind of want to co-late all the, co-locate the intelligence with the app
所以他们，他们有点想把这些都 co-late，把智能和 app co-locate 在一起
12:37
and the data as close as possible, right, that's what we've seen too.
而且数据要尽可能靠近，对吧，这也是我们看到的。
12:41
That's something that I didn't expect really because some of these, some of these responses
这真是我没预料到的，因为其中有些，有些回复
12:44
are kind of long, but it's like, we've seen with our customers at scale, like we want
有点长，但就像，我们在规模化客户那边看到的那样，我们想
12:48
to get the LLMs and the agents closer to us and our actual apps using them and the people
让 LLMs 和 agents 更靠近我们，以及我们实际使用它们的 apps，还有那些
12:52
using them as well. And not to mention, if you're using some of these SaaS providers, they're
也在使用它们的人。更不用说，如果你用一些 SaaS 提供商，它们
12:56
maybe not the most reliable in terms of what traditional enterprises expect in demand.
就传统企业对需求的要求而言，可能并不是最可靠的。
13:02
So you can't just have them, you know, going offline for hours because, you know, even,
所以你不能就让它们，你知道，一连离线好几个小时，因为，你知道，甚至，
13:06
even at my past job, the very first thing I did there was work on a warehouse inventory system,
甚至在我上一份工作，我在那儿做的第一件事就是做一个仓库库存系统，
13:11
and let's just say it was for a large retailer that was a grocery retailer in the US and I was
这么说吧，那是给美国一家大型零售商做的，一家食品杂货零售商，而我当时
13:17
in, you know, intern at a college and the app was older than me at the time, still older than me.
在大学里，你知道，做实习生，而且那个 app 当时比我还老，现在也还比我老。
13:21
And you know, it was just, it was a bunch of like, you know, you know,
而且你知道，它其实就是，就是一堆，像是，你知道，你知道，
13:26
C and shell scripts on unique systems. And if that thing went down, like within five minutes,
C 和 shell scripts，跑在一些很特别的系统上。如果那东西挂了，差不多五分钟之内，
13:31
like the warehouse was backing up trucks, like on the interstate. And so it's like, you have to
像是仓库那边就会让卡车堵到，像是，州际公路上。所以这就像，你必须
13:37
like, the scale of enterprise versus just, you know, how, you know, like consumer is a completely
就像，enterprise 的规模，对比只是，你知道，怎么说呢，consumer 的，完全是
13:42
different scale. So that's why some of all the things were trying to make them a little bit more
不同的规模。所以这就是为什么，所有这些事情里，有些我们正试着让它们更
13:47
enterprise grade and more, more used to what these enterprises expect from software.
enterprise grade，而且更，更符合这些 enterprises 对 software 的期待。
13:55
Some mornings as I'm drinking my morning coffee, I listen to live news updates to figure out what's
有些早上，我一边喝着早上的咖啡，一边听实时新闻更新，想弄清楚世界上发生了什么。
14:01
going on in the world. And it seems like recently everyone is talking about AI kill switches.
而且最近好像每个人都在聊 AI kill switch。
14:08
But the reality is that there is no one AI kill switch. We've moved from models to agents,
但现实是，并不存在某一个 AI kill switch。
14:14
those agents are connected to multiple models. They have an agent harness. They're connected to
我们已经从 models 转向了 agents，
14:19
MCP servers and tools. And those agents are acting within a fleet, delegating to one another and
这些 agents 连接着多个 models。
14:25
interacting with one another. In that environment, what you don't have is a single kill switch that
它们有一个 agent harness。
14:32
solves all of your problems. But that doesn't mean that you can't exert your control. You just have
它们连接着 MCP servers 和 tools。
14:38
to exert that control across various layers of that stack through, of course, component input,
而这些 agents 是在一个 fleet 里行动，彼此委派任务，彼此交互。
14:45
output, safeguards, but also controls on agents tied to their agent identity, things that track
output、safeguards，还有跟 agents 的 agent identity 绑定的 controls，以及那些用来追踪
14:53
tool misuse or memory poisoning or goal drift, etc. And what we're providing at prediction guard,
tool misuse、memory poisoning 或 goal drift 等等的东西。而我们在 prediction guard 提供的是，
15:01
the company that I lead personally, is an AI control plane that you self-host in your own
我个人带领的这家公司，是一个 AI control plane，你可以把它 self-host 在你自己
15:07
infrastructure that gives you that control plane or control layer for the relevant
的 infrastructure 里，为相关的
15:14
stack that agents operate on top of. I'd encourage you to check out what we're doing at prediction
stack 给你提供那个 control plane 或 control layer，让 agents 在其上运行。我鼓励你去看看我们在 prediction
15:20
guard.com slash practical AI. Go ahead and book a call with myself and the team to learn more about
guard.com/practical AI 上做的事。欢迎直接和我以及团队预约一个 call，进一步了解
15:27
how you can gain control of the agents operating in your environment and exert zero trust principles
你如何掌控在你环境里运行的 agents，并在你操作这些 agents 的方式中
15:36
in the way that you operate those agents. So check us out at prediction guard.com slash practical AI.
施加 zero trust principles。所以来 prediction guard.com/practical AI 看看我们。
15:43
That's prediction guard.com slash practical AI.
那就是 prediction guard.com slash practical AI。
15:49
So Nick, I would love to dive back. I want to kind of build on what you've, what you've been
所以 Nick，我很想再深入聊聊。我想顺着你刚才、你一直在
15:53
sharing so far. And part of that is that we're all diving into this multi-agentic world where
分享的内容继续讲。而这其中一部分是，我们都在进入这个 multi-agentic 世界，在那里
16:01
some late last year you'd have an agent. And then for most people I think getting into multi-agents,
大概去年晚些时候，你会有一个 agent。然后对大多数人来说，我觉得进入 multi-agents，
16:10
it was probably developers and the environments that the AI providers were making available
可能还是开发者，以及 AI providers 当时在提供的那些 environments
16:19
and you had multiple agents working on a project. But now we're really seeing with these work
然后你会有多个 agents 在一个 project 上工作。但现在，随着这些工作
16:24
products and other things that have come out. We're seeing agents really multiplying across
产品和其他已经出现的东西，我们真的看到了。我们看到 agents 正在
16:31
many different contexts and with a lot of different utility. And it's moved out of being
许多不同的 contexts 中成倍增加，并且有各种不同的用途。而且它已经不再只是
16:37
just frontier services. And we're now doing that with our hosted agents as well that we're running.
就是 frontier services。我们现在也在用自己正在跑的 hosted agents 做这件事。
16:44
And so like this multi-agentic world has just taken off in a very short amount of time.
所以这个 multi-agentic 世界在很短的时间里就一下子起飞了。
16:50
And as you're, I guess I'm asking if you can paint a picture as you start scaling this out
而随着你——我想我是在问，你能不能在你开始 scaling 这套东西的时候描绘一下画面
17:00
in the enterprise and what that looks like. Like could you describe,
在企业里会是什么样子。能不能描述一下，
17:05
we have analogies as you pointed out with the traditional software development world and the
我们有一些类比，就像你指出的，跟传统软件开发世界以及
17:10
containers and how you're structuring deployment and stuff. Could you talk a little bit about
containers，还有你是怎么组织 deployment 这些东西的。你能不能稍微讲一下
17:16
what that looks like, kind of paint a picture so that somebody can really get that in their head
那会是什么样子，稍微描绘一下，好让一个人真的能在脑子里理解这件事
17:23
me first of all about what that multi-agent world looks like when you were doing,
我首先想知道，当你当时在做的时候，那个 multi-agent 世界是什么样子，
17:29
when you were doing all these right things that you need to do in the enterprise.
当你在企业里做那些你必须做的、正确的事情时。
17:33
Yeah. So I mean, I think that the first thing that we want to focus on is that
对。所以我的意思是，我觉得我们首先要关注的是，
17:37
the enterprise needs to get whatever technology approved throughout their system.
企业需要让不管什么技术都在他们整个系统里获得批准。
17:43
And then we, so we are providing a harness called the Tainsuit Agent Build Pack. And that helps
然后我们，所以我们提供了一个叫 Tainsuit Agent Build Pack 的 harness。它能帮助
17:50
through that process. So our existing customers already have our customer, you know,
走完那个流程。所以我们的现有客户，你知道，已经是我们的客户了，
17:53
we're already approved so that we can help provide them those tools. And we just make that process
我们已经被批准了，所以我们可以帮他们提供那些工具。我们只是让那个流程
17:58
just another app to play it on the platform. So I probably haven't covered. They said,
变成 platform 上又一个能用的 app。所以我可能还没讲到。他们说，
18:03
I'll explain what I build pack is, right? So when you, when you, yeah, I was about to ask you that
我来解释一下 Build Pack 是什么，对吧？所以当你，当你，对，我正要问你那个
18:07
if you kind of dive into what the specifics of each of those things does.
如果你稍微深入了解一下这些东西每一个具体是做什么的。
18:11
Sure. So when I would do demos before AI and before, you know,
当然。所以以前我做 demo 的时候，在 AI 之前，还有，你知道，
18:15
a year or two ago or whatever, we would always take like a Java app and it was like a simple,
一两年前或者什么的，我们总是会拿一个 Java app，而且它就像一个简单的，
18:19
like kind of like, you know, music album called it spring music or whatever. But we take that Java
有点像，你知道，音乐专辑，叫它 spring music 或者什么的。但我们会把那个 Java
18:24
code to be like a jar file, like compile Java code and we're like, all right, now we're going to take
代码弄成一个 jar file，就像 compile Java code，然后我们说，好吧，现在我们要拿
18:27
this. We have a little simple manifest file that tells us what the, what the name of the app is,
这个。我们有一个小小的简单 manifest file，它会告诉我们，那个，app 的名字是什么，
18:31
how many, how much resources to give it. And we just run a command they called see a push.
要给它多少、多少资源。然后我们就运行一个他们叫作 see a push 的命令。
18:35
And that would then just take that code, the jar file and upload it to platform. And the platform
然后它就会直接把那个代码、那个 jar file 上传到 platform。然后这个 platform
18:39
would be like, Oh, it's a Java app. So I'm going to use the Java build pack and the build pack is,
就会像是，哦，这是个 Java app。所以我要用 Java build pack，而 build pack 就是，
18:44
it's basically a kind of a command set to make a best practice container based off that use case.
它基本上就是一种 command set，用来基于那个 use case 做出一个 best practice container。
18:50
So the Java build pack knows to do all the JVM memory calculators, do the JDK and the certificates
所以 Java build pack 知道怎么处理所有的 JVM memory calculators，处理 JDK 和 certificates，
18:56
and all everything that's best practice around Java and the container on a cloud platform.
以及所有那些围绕 Java 和 cloud platform 上的 container 的 best practice。
19:01
And that would be great. And then we would attach database and you know, all these things and show
那就会很棒。然后我们会 attach 上 database，还有你知道的，所有这些，然后展示
19:06
how that that scales. Now we've taken that same concept and then said, Hey, we're going to have
它怎么 scale。现在我们把同样的概念拿过来，然后说，嘿，我们要有
19:10
this agent, we're going to just see a push an agent. And instead of like, you know, machine readable
这个 agent，我们就是要，你看，push 一个 agent。然后不像，你知道，machine readable
19:15
code, we have a human readable language, the agent, some D. So basically tell what the agent,
code，我们有 human readable language，这个 agent，some D。所以基本上就是告诉这个 agent，
19:21
it's going to do. And then you just push it to the platform and it starts up with, you know,
它要做的事。然后你只要把它推到 platform 上，它就会启动，你知道，
19:25
within a minute type of thing. And you can scale that up and you push as many agents as you want.
一分钟之内那种。而且你可以把它 scale up，想推多少 agents 就推多少。
19:29
And I think the coolest thing that I've seen is that once we're on the platform, then we can run
而且我觉得我见过最酷的事情是，一旦我们上了 platform，我们就可以跑
19:33
models on the platform. So you know, obviously a lot of our customers are airgapped. And some of them
platform 上的 models。所以你知道，显然我们很多客户都是 airgapped。而且有些客户
19:38
were, we're smart enough to buy GPUs a few years ago. So some of them have lots of GPUs that can run
几年前就够聪明，买了 GPUs。所以他们有些有大量的 GPUs，可以跑
19:44
all these local models and have a great experience. Some are wishing that they bought GPUs and, you know,
所有这些 local models，而且有很棒的体验。有些就在想，当初要是买了 GPUs 就好了，而且，你知道，
19:48
hardware as the pricing of all this hardware goes through the roof. But then they can run
hardware，因为所有这些 hardware 的价格都涨上天了。但那样他们就可以跑
19:54
local models and tie the agent to that. Or if they don't have, you know, they might have a contract
local models，然后把 agent 绑定到那上面。或者如果他们没，你知道，他们可能有一份合同
19:58
with one of their cloud providers or one of their, you know, one of their approved LMS the choice.
用他们的某个 cloud provider，或者他们其中一个，你知道，他们认可的某个 LMS，作为选择。
20:03
And we can register that. Then that kind of gave us a simple like, here's your out-of-box agent
然后我们可以把那个注册上。那有点像给了我们一个简单的，就像，这是你的开箱即用 agent
20:08
with chat experience. And then, you know, along the way, like, we've done a lot of things with MCP.
带 chat 体验。然后，你知道，一路下来，就像，我们用 MCP 做了很多事情。
20:13
So you know, the model context protocol, I probably you guys covered that on the show before,
所以你知道，model context protocol，我估计你们之前在节目里聊过这个，
20:18
but it's, you know, basically way for an app to talk to an LLM, like very, very high level, right?
但它其实，你知道，基本上就是让 app 跟 LLM 对话的一种方式，非常非常 high level，对吧？
20:22
So we, that's become, you know, useful as well as we then give the agent tools via MCP servers.
所以我们，那变得，你知道，挺有用的，而且我们之后通过 MCP servers 给 agent 工具。
20:28
And even at Broadcom, like, the only way approved way to use for our developers is to deploy MCP
而且就算在 Broadcom，像，对我们开发者来说，唯一被批准的使用方式就是部署 MCP
20:35
servers on on Tazy Platform. So that we have a whole set of MCP servers that are approved and
servers 在 Tazy Platform 上。这样我们就有一整套被批准的 MCP servers，而且
20:40
hosted and they're kind of secured and maintained. Because they are very similar to applications as
被托管，而且它们算是被保护和维护着。因为它们和 applications 非常相似，就像
20:44
well. And then we have a concept of like the MCP gateway service that kind of registers that and
一样。然后我们有这么一个概念，像 MCP gateway service，它会注册那个，并且
20:49
cures all those servers. So then we can then extend the gateway to that agent. So then the gateway,
管理好所有那些 servers。然后我们就能把 gateway 扩展到那个 agent。然后这个 gateway，
20:54
this agent now has like secure access to a runtime, you know, battle test runtime. It can
这个 agent 现在就像有了对 runtime 的 secure access，你知道，battle test runtime。它可以
20:59
talk to an LLM that's approved and get tools via and gateway. And then we're starting to add in like
跟一个经过批准的 LLM 对话，并通过 gateway 获取 tools。然后我们开始加入像
21:06
the memory service as well to have a kind of a showed memory service. So we're kind of expanding
memory service 也一样，来拥有一种有点像 showed memory service 的东西。所以我们有点像在扩展
21:10
that all that out where we're seeing all those best practice app lifecycle go to, you know, see a push,
把这一切都铺开，我们看到所有那些 best practice app lifecycle 都走向，你知道，看到一次 push，
21:17
go. Here's an agent here. Here's all the things it wants to do. And then, you know, it can be a
继续。这里有一个agent。这里是它想做的所有事情。然后，你知道，它可以是一个
21:21
it can be a security review agent. Like one of the things I showed I was at a
它可以是一个 security review agent。就像我展示的其中一个东西，我当时在一个
21:25
I think a week or two ago, I can't quite remember. I was at as in Las Vegas for our conference via
我想大概一两周前吧，我记不太清了。我当时在，也就是，Las Vegas 参加我们的会议，通过
21:29
our explorer. And I had like a five phase demo where I walk through this. Maybe I'll show a little
我们的 explorer。然后我有一个大概五个阶段的 demo，我在里面演示了这个。也许我会展示一点
21:33
bit of this at the Midwest summit here. But you know, push the agent bind bind all these services.
这个在 Midwest summit 上。但是你知道，push 这个 agent，bind bind 所有这些服务。
21:39
But the end the end goal was like the the boss level was like, all right. Now we have these agents
但是最终最终的目标就像是，那个 boss 关就像是，好吧。现在我们有了这些 agents
21:44
sitting in there. And I can hook up hook them up to a repo with like an event listener or a
坐在那里。然后我可以把它们 hook up 到一个 repo，用一个 event listener 或者一个
21:48
web hook and say, all right, now do a like, you know, issue five on my repo. Do you a security
web hook 然后说，好吧，现在做一个，就像，你知道，在我 repo 上的 issue 5。做一个 security
21:53
review on this repository? And then like it sends it off to the agent via web hook. And like, oh,
review 在这个 repository 上？然后就像它通过 web hook 把它发给这个 agent。然后就像，哦，
21:59
the Tainzoo agent like it just fires up gets it, you know, fires up from the platform. It gets a
Tainzoo agent 就像，它直接启动起来，拿到，你懂的，从平台上启动。它会收到一个
22:04
request. And it's just immediately pulls down the repo and starts doing a security review. And
request。然后它马上把 repo 拉下来，开始做 security review。然后
22:07
then publishes or pushes up to the issue that I created on GitHub and says, this is security review.
接着发布或者推送到我在 GitHub 上创建的那个 issue，然后说，这是 security review。
22:13
So you could see how that flow would be really useful in enterprise where like maybe on every
所以你能看出来，这个流程在企业里会非常有用，比如说可能在每一次
22:17
git commit, you just have all these agents spinning up on the fly that could be a security review agent.
git commit 时，你就能让所有这些 agent 即时启动起来，它可以是 security review agent。
22:22
You know, potentially like a, you know, a style or, you know, code quality review agent or
你懂的，可能像，你懂的，一个 style 或者，你懂的，code quality review agent，或者
22:28
potentially even like a, I don't know, maybe a design review agent all firing up on the fly as
甚至可能像，我不知道，也许一个 design review agent，全都随着
22:34
things like call. So it's kind of like, and then they all have their they're all in their own
像 call 这样的事情即时启动。所以就有点像，然后它们都有各自的，它们都在自己的
22:38
little sandbox and everything's proper, right? Because part of, you know, the platform,
小小的 sandbox，而且一切都妥妥的，对吧？因为，你知道，平台的一部分，
22:42
for a while, it's been like this very secured container runtime where like we have the, you know,
有一阵子，它一直是这种非常安全的 container runtime，就像我们有，你知道，
22:46
they can't talk to each other unless you've enabled them so they can't go to the internet
它们不能互相通信，除非你启用了它们，所以它们不能上 internet
22:49
unless it's enabled they can't, you know, escalate otherwise they'll only get access to what you
除非它被启用，否则它们不能，你知道，escalate，不然它们只能访问你
22:53
give them. So it's kind of that benefit as well. And then obviously they're containerized,
给它们的东西。所以这也算一种好处。然后显然，它们是被 containerized 的，
22:58
they don't have access to the underlying host system as well. So the, you know, as we see,
它们也无法访问底层的 host system。所以，你知道，正如我们看到的，
23:04
you know, some of these things of uncontrolled agents forms, I think we have some remedies here
你知道，这些 uncontrolled agents 形成的一些情况，我觉得我们这里有一些补救办法
23:08
that some of these, some of these AI labs can maybe take into account and then implement some
其中一些，其中一些 AI labs 也许可以考虑进去，然后实施一些
23:13
best practices that's went throughout the past two decades. The unplanned forms, the forms
过去二十年里一直沿用的 best practices。那些计划之外的形式，那些
23:19
that you don't necessarily want. We actually have a two-second follow-up. And that is it's for my
你未必想要的形式。我们其实还有一个两秒钟的 follow-up。而且这是为了我的
23:25
benefit. And you know, we've talked a lot on the show over various episodes about agents, you know,
好处。而且你知道，我们在节目里不同集数里聊了很多关于 agents 的内容，你知道，
23:31
using MCP to get access to different apps and services and tools and stuff like that out there.
用 MCP 去访问不同的 apps、services、tools 以及外面那些类似的东西。
23:37
We haven't really talked about what an MCP gateway is. And I was wondering if you could share
我们其实还没真正聊过 MCP gateway 是什么。我在想你能不能分享一下
23:42
that for a moment just as a detail because that's kind of the one little new piece of architecture
这个，就一会儿，作为一个细节，因为那算是那个小小的新架构
23:46
that we've never touched on on the show. Sure. Yeah, so if you think about an MCP gateway as a way
我们在节目里从没碰过的。当然。是的，所以如果你把 MCP gateway 看作一种方式
23:51
to easily control and scale access to MCP servers, especially within a year in the cloud, I guess
能轻松控制和 scale 对 MCP servers 的访问，尤其是在 cloud 里的一年内，我猜
23:59
platform. So we may have like 30 or 40 MCP servers that we run or I don't know how many it is, right?
platform。所以我们现在可能跑了大概 30 或 40 个 MCP servers，或者我也不知道到底有多少个，对吧？
24:06
And then those get bound, those get registered and bound to certain specific MCP gateway. So maybe
然后这些会被绑定，会被注册并绑定到某个特定的 MCP gateway。所以也许
24:12
there's a cut, you know, and the MCP gateway is something we can spit up on the fly on the platform.
有一个分界，你懂的，而 MCP gateway 是那种我们可以在 platform 上临时快速弄出来的东西。
24:17
There are many different MCP gateways out there. There's many different AI gateways out there.
外面有很多不同的 MCP gateways。也有很多不同的 AI gateways。
24:21
Some of them are like the same, one and the same. But you basically register MCP servers to this gateway
有些其实是一样的，同一个东西。但你基本上就是把 MCP servers 注册到这个 gateway
24:27
and the gateway kind of controls access, so to speak and regulates that. So it may say like you can
然后这个 gateway 某种程度上控制访问，可以这么说，而且会做监管。所以它可能会说，比如你可以
24:33
bind to this gateway and you only get these MCP servers, you only get these MCP tools within it,
绑定到这个 gateway，然后你只能拿到这些 MCP servers，你只能拿到它里面的这些 MCP tools，
24:39
that type of thing. So it's a way to easily abstract all the different MCP servers that may be an
那种东西。所以这是一种很方便地抽象掉所有不同 MCP servers 的方式，可能会是一个
24:43
organization and kind of direct them into a certain, either a certain agent or we even have a flow
组织，然后有点把它们引导到某个特定的，要么是某个特定的 agent，要么我们甚至有一套 flow
24:49
where like you can, we'll take those MCP gateways that we run that front all of our MCP servers.
就是你可以，我们会拿那些我们运行的 MCP gateways，它们作为我们所有 MCP servers 的前端。
24:54
And then you can say, hey, go give me a way to connect this to my cursor or my cloud code or my
然后你可以说，嘿，给我一个办法，把这个连接到我的 cursor，或者我的 cloud code，或者我的
24:59
whatever. And then it just injects it in there. And then your local agent has all the tools that are
随便什么。然后它就直接把它注入进去。然后你的本地 agent 就拥有所有那些
25:04
approved as well for your organization. And then it's just like a simple click and then it just
也为你的组织批准过的工具。然后就像简单点一下，然后它就直接
25:10
registers. And then like depending on the MCP server helps the panel the office well because some of
注册。然后呢，取决于 MCP server，它也能很好地帮助 panel the office，因为其中有些
25:15
them are like, well, if it's my, if I'm using the GitHub MCP server, I want to pass my end user
它们会想，好吧，如果是我的，如果我在用 GitHub MCP server，我想把我的 end user
25:20
credentials all the way through to that MCP server because I don't want, you know, I don't want it
credentials 一路传到那个 MCP server，因为我不想，你知道，我不想要它
25:25
just to be like, well, here's our general generic service count for the entire organization. And
就像，嗯，这是我们整个组织总体的通用 service count。而且
25:28
every GitHub action is going to be seen by that one account. And you have to have the individual
每个 GitHub action 都会被那一个账号看到。而且你得让个人的
25:34
identity pass through. So you have like the concept of like SSO and sign in as well for an
identity 透传过去。所以你得有类似 SSO 和 sign in 的概念，也要给一个
25:39
identity for agents or humans running their agents on their desktop. So that's kind of a quick
identity，给 agents 或者在自己 desktop 上跑 agents 的人类。所以这算是个快速的
25:44
level view. And then a great way about that too is because it's all flowing to that gateway,
概览。然后关于这点还有个很棒的地方，因为这一切都流向那个 gateway，
25:48
we get metrics out of it. So we can see all the toolcalls and all the all the events happening
我们能从中拿到 metrics。所以我们可以看到所有 toolcalls 和所有所有正在发生的 events
25:52
coming from the agents to the MCP servers and back and forth. So you kind of kind of see what's
从 agents 到 MCP servers 来回流动。所以你大概，大概能看到
25:57
happening, especially if maybe if there's something unexpected or like, well, this one agent,
发生了什么，尤其是如果可能有什么意外，或者像是，嗯，这个 agent，
26:01
you know, made 200,000 toolcalls to, you know, delete repo, like, oh, like, well, maybe we should
你知道吧，做了 200,000 次 toolcalls，就是，去删除 repo，就，哦，就，嗯，也许我们应该
26:07
alert on that. Like that's probably not good. Yeah. So, and you know, another visibility or metric
对那个发 alert。这大概不太好。对。所以，而且你知道，另一个 visibility 或 metric
26:13
to see what agents are doing, you know, what toolcalls are doing as well. So I guess, hopefully that
来看到 agents 在做什么，你知道，也看到 toolcalls 在做什么。所以我想，希望这
26:17
made sense because that was, that's kind of how I understand it. Right. Yeah, that was great.
说得通，因为那，那差不多就是我的理解。对吧。对，这很棒。
26:22
And there were a couple of things that you mentioned in there. I think are additionally worth
而且你刚才在里面提到了几件事。我觉得还额外值得
26:26
digging into. And you, you kind of have this, if I'm understanding right, in part of the,
深入挖一挖。而且你，你大概有这个，如果我理解得没错，在你正在做的，
26:33
the build pack that you're working on, part of that centers around kind of pushing an agent.MD
那个 build pack 里，其中一部分核心大概就是推送一个 agent.MD
26:39
file. You mentioned like people have these MD files hanging around. You mentioned memory,
文件。你提到说，人们身边就散落着这些 MD 文件。你提到了 memory，
26:44
you mentioned identity. I'm wondering if you could kind of help us understand like if I have,
你提到了 identity。我在想，你能不能稍微帮我们理解一下，比如说如果我有，
26:54
if I'm working on my laptop, right? And I'm using a certain agent harness. I may have an MD
如果我在自己的笔记本电脑上工作，对吧？而且我在用某个 agent harness。我可能会有一个 MD
26:59
file that is kind of generic. And then I have files that are created by the agent that it uses that
file，算是比较通用的。然后我还有一些由 agent 创建的文件，它用的那些文件
27:08
are only corresponding to my single agent. And then I have, you know, memory and then, you know,
只对应我这一个 agent。然后我还有，你知道的，memory，然后，你知道的，
27:14
calls into other things as you're pushing this like off the laptop. Now I imagine taking that off.
calls 到其他东西里，当你把这东西从笔记本电脑上往外推的时候。现在我想象把它拿出去。
27:20
Like what belongs in like an agents.md file. And what is, like as you're giving that example,
比如说，什么该放在一个 agents.md file 里。那什么是，就像你给那个例子时，
27:29
right, you may have one agent in the platform that is a security reviewer for,
对吧，你可能在 platform 里有一个 agent，是个 security reviewer，负责
27:34
or that does security reviews, right? When you spend that agent up, does each instance of that,
或者说做 security reviews，对吧？当你把那个 agent 启动起来的时候，它的每个 instance 是不是，
27:41
each session have a unique identity. What is the, what's unique about that session and the data that
每个 session 都有一个唯一的 identity。那这个，那个 session 的独特之处是什么，还有那些数据
27:50
that is generated by that session versus maybe the things that are always static? Like I'm not sure
是由那个 session 生成的，和那些可能一直 static 的东西相比呢？比如我不确定
27:55
if the agent's.md is always static. So could you help us understand like which are the static elements?
agent's.md 是不是一直 static。所以你能不能帮我们理解一下，哪些是 static 的元素？
28:00
What's developing over time? How is that like state of the agent spread across, you know,
什么会随着时间发展？agent 的 state 是怎么分布在，你知道，
28:07
the implementation, I guess? Sure. Yeah. So if you think about it, like in this specific
implementation 里的，我猜？当然。对。所以你想想，在这个具体的
28:12
implementation, the agent's build pack is basically like the boilerplate code of what you want the
implementation 里，agent 的 build pack 基本上就是你想要
28:18
agent or app to do. So like I'll do demos and I'll have to be like your pirate. Like like one of my
agent 或 app 去做什么的 boilerplate code。所以我做 demo 的时候，就得扮演你的海盗。就比如我的一个
28:23
demos is like you're a, you're a, you're a, a, Jira, like you're a senior software engineer.
demo 是，你是一个，你是一个，你是一个，一个，Jira，你是一个 senior software engineer。
28:30
You know, you're gonna, you're gonna watch this Jira like a page or I forgot the terminology here.
你知道吧，你会，你会盯着这个 Jira，就像盯一个页面一样，还是我忘了这里的术语了。
28:36
But basically watch this Jira queue and look for any inbound tickets as inbound tickets come in,
但基本上就是盯着这个 Jira queue，看看有没有 inbound tickets，只要 inbound tickets 进来，
28:41
review them and like, you know, make sure that they're a good quality. Like here's all these
就审阅它们，然后，你知道吧，确保它们质量不错。就像这里有所有这些
28:45
different things that you want to do and like don't do any of this. So it's more of a very simple
不同的、你想做的事情，然后像“这些全都别做”。所以它更像是一个非常简单的
28:48
agent, something where like it's just telling like this agent what, what it should do and how it should
agent，就是那种，你只是告诉这个 agent 它该做什么，以及它该怎么
28:54
act. And then you, you can call that and how like, all right, well, like, you know, ticket 500 came in,
行动。然后你，你可以调用它，然后就像，好吧，嗯，你知道吧，ticket 500 进来了，
28:59
it was poorly documented and like update that Jira ticket, you know, whatever. And then you can add
它的文档写得很差，然后就像更新那个 Jira ticket，你懂的，随便吧。然后你可以加
29:03
some things like talk like a pirate. So it's like, oh, my tea, your, you know, ticket 500 is a
一些东西，比如像海盗一样说话。所以就像，哦，我的茶，你的，你知道吧，ticket 500 是一个
29:07
terrible, right? Like go back and, you know, update that type of thing. So that's kind of where
很糟糕，对吧？就像回过头去，你知道，更新那类东西。所以这大概就是
29:11
the static part of it is. And then like the memory service we have, you can attach it, you can have,
它 static 的部分所在。然后就像我们有的 memory service，你可以把它接上去，你可以有，
29:16
you know, you know, potentially memory of a certain project. So, you know, we kind of envision,
你知道，你知道，可能是某个项目的 memory。所以，你知道，我们有点设想，
29:21
like each set of teams to have their own multiple set of agents, right? And it could be like, well,
比如每一组团队都有自己多组 agents，对吧？而且可能是，嗯，
29:25
you know, team A's agents have the access to team A's memory. So as they come up, they know, like
你知道，team A 的 agents 可以访问 team A 的 memory。所以当它们启动时，它们就知道，就像
29:31
all of this historical data of like what, like the architecture of the application, the, you know,
所有这些 historical data，比如什么，比如 application 的 architecture，那个，你知道，
29:36
what's been done before, you know, like the roadmap, that type of thing. So as it spins up and does
之前做过什么，你知道，比如 roadmap，那类东西。所以当它启动并做
29:41
like a security review, it can make, you know, proper decisions and proper guidance. So that, you
比如 security review，它就能做出，你知道，合适的决定和合适的指导。所以这样你
29:47
know, as you kind of build your local agents and kind of give them memory and they become smarter,
你知道，当你逐渐构建自己的 local agents，并给它们 memory，它们就会变得更聪明，
29:51
that memory service as you have agents spin up and down, they just immediately, they immediately know
那个 memory service，当你的 agents 不断 spin up and down 时，它们就会立刻，马上就知道
29:56
what's happened before, know how to proceed. As you heard at the beginning of this episode,
之前发生了什么，知道该怎么继续。就像你在本期节目开头听到的那样，
30:05
some amazing people like Nick, who's the guest on this episode are joining us at the Midwest
一些很厉害的人，比如 Nick，也就是本期的嘉宾，会和我们一起参加 Midwest
30:11
AI Summit, October 15th in Indianapolis, Indiana. Nick is going to come talk about what it looks like
AI Summit，10月15日在 Indianapolis, Indiana。Nick 会来聊聊，在真实世界环境中 deploy AI agents
30:19
to deploy AI agents in real-world environments. And we've got a bunch of other amazing speakers
会是什么样。我们还有一群其他很棒的讲者
30:26
talking about how AI enabled teams work, how to make decisions about what to say yes to and what
聊 AI enabled 团队如何运作，如何决定该对什么说 yes，以及
30:33
to say no to in relation to AI and much more super practical and amazing talks. Plus, there's going to be
该对什么说 no，尤其是在 AI 相关的事情上，还有更多超实用、超精彩的分享。而且，还会有
30:41
an AI engineering lounge where you can actually sit down with folks like myself, I'll be there,
一个 AI engineering lounge，你可以真正坐下来和像我这样的人聊聊，我也会到场，
30:47
get, get review on your AI architecture, your stack, your roadmap, etc. towards the rollout of things
拿到、拿到关于你的 AI architecture、你的 stack、你的 roadmap 等等的 review，为了推进
30:56
like agents or other AI technology. This is a practical, amazing event that you don't want to miss.
像 agents 或其他 AI 技术这类东西的 rollout。这是一个很实用、超棒的活动，你绝对不想错过。
31:03
There'll be great food, great people, great networking and practical value. Do not miss this,
会有很棒的食物、很棒的人、很棒的 networking 和实际价值。千万别错过，
31:10
it's October 15th in Indianapolis, Indiana. And you can use the code practical AI20 to get 20%
时间是 10 月 15 日，在 Indianapolis, Indiana。你可以用优惠码 practical AI20 获得 20%
31:19
off registration. So go to MidwestAISummit.com and use the code practical AI20 to get 20% off registration.
的注册费折扣。所以去 MidwestAISummit.com，用优惠码 practical AI20 获得注册费 20% 的折扣。
31:30
Before we were going into break, I think I may have cut you off when you were just starting to say
在我们进入休息之前，我想我可能在你刚开始要说点什么的时候打断你了，
31:38
something you want to dive back into that real quick. No, I think I covered it up. Hopefully I answered
你想不想马上再回到那个话题？没有，我想我已经说完了。希望我回答了
31:45
the question again, but that's no problem. So really, this has been really good for helping me
又是这个问题，不过没关系。所以真的，这对我帮助很大，
31:52
trying to conceptualize how all this fits together. I think I had some understanding, but
帮我去想清楚这一切是怎么拼在一起的。我觉得我本来是有一些理解的，但
31:59
you've gotten a level of detail. I want to ask you as we are looking forward, I have several questions
你已经讲到了这种细节程度。我想在我们往前看的时候问你，我有几个问题
32:08
that are really kind of driving my thinking. You've got me thinking creatively based on what you've
这些问题真的有点在推动我的思考。你让我基于你刚刚
32:13
said. So as you look at, we've talked about MCP, there are new things coming out. Things like,
说的内容开始创造性地思考。所以你看，我们聊过 MCP，现在有新东西出来。比如，
32:20
there's the A2A which is the agent to agent protocol and organizations are starting to look at
有个 A2A，也就是 agent to agent protocol，组织们开始关注
32:26
that and they already have their MCP servers maybe, but they're trying to get the structure and
它，而他们可能已经有自己的 MCP servers 了，但他们想搞清楚结构，
32:35
kind of bring the sanity around it that you've been talking about. Then they kind of go,
并围绕它把你一直在说的那种条理性带进来。然后他们就会说，
32:40
oh, okay, it's not so different from what we've already know. But as this is evolving very fast
哦，好吧，这跟我们已知的其实没那么不一样。但因为这个东西发展得非常快
32:46
and you have things like these new protocols, do you have any guidance on how to start integrating
而且现在还有这些新的 protocols 之类的东西，你有没有什么指导，告诉我们该怎么开始整合
32:51
this constant flow of change in in terms of, first on the technology, I'll hit the other half of
这种持续不断的变化，具体来说，先谈技术这一块，另一半我之后会讲到
32:57
that afterwards. As you're trying to bring things in and make it work, have you guys thought
那部分之后。当你们试着把这些东西引进来并让它跑起来时，你们有没有想过
33:05
about how to structure that? I think it's, I mean, I guess we'll just say, take baby steps
该怎么组织这个？我觉得，我是说，我猜我们就只能说，小步慢走
33:11
from everything, because especially when you talk about the enterprise scope, they can be a lot
每件事都这样，因为特别是说到 enterprise scope 的时候，它们可能会慢很多
33:17
slower to adopt some of these newer technologies. So you kind of layer that on and like,
才去采用这些较新的技术。所以你就是一点点叠加上去，然后就像，
33:22
you don't have to necessarily adopt all of these things at once. You can maybe just get a simple agent
你不一定非得一次性采用所有这些。你也许可以先搞一个简单的 agent
33:27
running. And for most, those places, just getting access to an LLM is still the biggest hurdle,
正在运行。而且对大多数那些地方来说，光是能访问到一个 LLM 仍然是最大的障碍，
33:35
right? A secured approved. And then potentially, going through the AI council,
对吧？先拿到安全审批。然后可能还要过 AI 理事会，
33:41
or the committee, everything must run through the AI council will be approved. If you're not
或者委员会，所有东西都必须经过 AI 理事会才能获批。如果你不是
33:47
if you've never worked an enterprise, most large enterprises have a capital or committee.
如果你从没在企业工作过，大多数大型企业都有一个资本委员会。
33:51
So I know for everything. It's like, well, we shall go to the committee and we shall present our
所以我知道，所有事情都是这样。就像，好吧，我们要去委员会，我们要展示我们的
33:57
idea and we'll wait six months and see what has the result been, right? So just kind of have to
想法，然后等六个月，看看结果如何，对吧？所以就只能
34:01
work through the councils and the committees and all that fun. But once you finally get something
走完这些理事会和委员会，还有所有这些好玩的事。但一旦你终于让某件事
34:05
approved, you just just iterate and you like, don't try to boil the ocean, make baby steps
获批了，你就只管 iterate，就像，别想一口吃成胖子，要小步前进
34:11
and get progress. And then as you'll, like, and you'll learn the technologies in the next step,
然后取得进展。接着你就会，怎么说呢，你会在下一步学到这些技术，
34:16
we'll just kind of be obvious, like, well, I can make the agent better if I get some MCP,
那就会挺明显，就像，嗯，如果我搞到一些 MCP，我就能让 agent 变得更好，
34:20
so just get them tools. All right, and then we've got to get these tools approved and then add on
所以直接给他们 tools 就好。好，然后我们得让这些 tools 通过审批，然后再往上加
34:24
to that. And then, you know, and some of these guardrails help, you know, of Tains and Platform
到那上面。然后，你知道，这些 guardrails 里有些，你知道，是 Tains 和 Platform 的，能
34:30
help enterprises bring these tools to light, because it's not just, you know, all this random stuff
帮助企业把这些 tools 用起来，因为这不只是，你知道，所有这些乱七八糟的东西
34:36
downloaded from the internet, right? Which that can be its own danger nowadays. So having the,
从 internet 上下载来的，对吧？这在如今本身就可能是一种危险。所以，
34:42
you know, having some some guardrails in place is a good thing. If you've had, you know,
你知道，有一些 guardrails 到位是件好事。如果你曾经有过，你知道，
34:47
if you have experience with some of these platforms or apps running before, take that,
如果你以前有运行过其中一些 platforms 或 apps 的经验，那就把那个，
34:52
take that experience and that knowledge and kind of bring it forward and utilize your existing
把那些经验和那些知识带过来，某种程度上把它们往前推进，并利用你现有的
34:57
people and skills, I think too. I think a lot a lot of times do I see an organization like a new
人员和技能，我也这么觉得。我觉得很多很多次，我会看到某个组织，就像有一个新
35:01
technology will come up and like a new team will be built and that new team will decree, but there's
技术会出现，然后像是会组建一个新团队，而那个新团队会发号施令，但是有
35:05
so much other like really skilled people that could help. And it's always kind of a friction too,
那么多其他的人，真的很有技能，本来能帮忙。而且这总是有点像摩擦，
35:10
so how try to try to get the teams to work together too and not just have the new team doing
所以怎么试着去、试着去让这些团队也一起协作，而不只是让新团队去做
35:14
all the new things and then the team that's the teams that been there all the while have that,
所有新的事情，然后那些团队，那些一直都在那儿的团队有那些，
35:18
like that's always another constant flow. So try to try to break the silos down as well to help,
就像那总是另一股持续不断的流动。所以也试着去、试着去打破这些 silos，来帮助，
35:23
you know, get people up to speed and and get through that process. Having, having worked in multiple
你知道，让大家尽快跟上，并且、并且把那个过程走完。有过、有过在多个
35:28
enterprises myself, I think, you know, that last point, I think I've seen that multiple times
企业里，我自己觉得，你知道，最后那一点，我觉得我在不同的组织里见过很多次
35:35
across different organizations and, you know, how do you bridge that? I think part of it, you know,
而且，你知道，你怎么去弥合这个？我觉得一部分是，你知道，你有了，每个组织都有自己的政治，它们都有不同的结构。
35:41
you have, you kind of have every organization has its own politics, they all have kind of the
很常见的是说新技术，新团队，但后来我觉得人们会认识到，你知道，为什么？你知道，那种，你会看到这种情况，有时候你会看到一种调和，但每次发生的时候似乎总是有点挣扎
35:46
different structures. It's really common to say new tech, new team there, but then I think people
去试着，你知道，你对此有什么想法吗，当你试着去做的时候，你知道，你可能会有，你知道，很多大组织都有，你知道，它们非常像一个，一个业务
35:53
recognize, you know, why? You know, that kind of, you'll see that and sometimes you'll see
识别，你知道，为什么？你知道，那种，你会看到那个，有时候你会看到
35:58
a reconciliation, but it always seems like a little bit of a struggle when that happens
一种对账，但每次发生这种情况时，似乎总是有点挣扎，
36:04
to try to, you know, do you have any thoughts around as you're trying to do that, you know, you
试图，你知道，当你试图这么做的时候，你有什么想法吗，你知道，你
36:11
may have, you know, a lot of big organizations have, you know, they're very like one, one business
可能有很多，你知道，很多大组织都有，你知道，它们非常像是某一个、某一个业务
36:18
unit, maybe almost like a different universe from another business unit sometimes. And so how do you
业务单元，有时候可能几乎像是另一个业务单元所在的平行宇宙。那你要怎么
36:25
get that kind of cross adoption where you're taking advantage of maybe both the new teams that
获得那种 cross adoption，让你既能利用可能存在的新团队，
36:33
may be there, but also some of the old structure teams and you're trying to spread the capability
也能利用一些旧结构里的团队，同时你还在试着把这种能力
36:39
across the larger organization in a useful way. Do you have any guidance? You know, I imagine
以有用的方式推广到更大的组织里。你有什么建议吗？你知道，我想象
36:46
it can't be that different from, you know, the more traditional app dev side, but I think
这不会和，你知道，更传统的 app dev 那边有太大不同，但我觉得
36:51
I think this is a struggle that a lot of that a lot of words are facing right now. Yeah, well,
我觉得这是很多、很多 words 现在正面临的一个难题。对，嗯，
36:56
I think, yeah, so I mean, it goes, I think a lot both ways. So if you have the new team, right?
我觉得，对，所以我的意思是，这很大程度上是双向的。所以如果你有新团队，对吧？
37:01
The new team can like the responsibility of the new team is like engage every other team because
新团队可以，像是，新团队的责任就是去接触其他每一个团队，因为
37:06
they have a lot of people that can help you out and achieve your goal, right? Because the new team
他们有很多人能帮你、帮你达成目标，对吧？因为那个新团队
37:09
probably most likely has a very stressful and like, you know, like adopt AI everywhere or before
很可能压力特别大，而且就是，你知道，像要到处都上 AI，或者以前
37:16
it was like go cloud everywhere, right? Like those those goals are like not quite easy to get going
就像到处都上 cloud，对吧？就那些目标其实不太容易推动起来
37:22
in an enterprise so that the new team needs to like, you know, potentially run, you know, office hours,
在 enterprise 里，所以新团队就得，就是，你知道，可能得搞，你知道，office hours，
37:28
weekly office hours, reach out to their other teams, you know, maybe before, you know, like we
每周的 office hours，去联系其他团队，你知道，也许之前，你知道，就像我们
37:33
like a lunch and learn if you're in the office and have people in the office that type of thing.
就像 lunch and learn，如果你在办公室，而且办公室有人，那种事情。
37:38
And just have like the forum for people to come in and learn and get started and like ask questions
然后就是有个 forum，让大家进来学、上手，然后问问题
37:43
and not, you know, like, hey, like I I saw this, you know, this post about, you know, this new
而不是，你知道，像，嘿，我我看到了这个，你知道，这个帖子，关于，你知道，这个新的……
37:48
platform or this new AI thing. How do I get started? Is there any documentation or what can I do to
平台，或者这个新的AI东西。我该怎么开始？有没有什么文档，或者我能做什么来
37:53
get started? Just kind of have like an open and like, you know, very welcoming place to get
开始？就像是有一种开放的、你知道，非常欢迎的地方，可以
37:58
started to bring in the folks. And then if you're on some of the, you know, let's say not new teams
开始，把人们带进来。然后如果你在一些，你知道，比如说不是新团队的地方
38:03
or teams that then are while being gauged in and reach out to those teams because a lot of the
或者那些团队，然后他们在被 gauged in 的同时，会去联系那些团队，因为很多
38:08
the new people are going to need help. Like they, you know, they're maybe under stress. So it's just
新人会需要帮助。就像，你知道，他们可能压力很大。所以这只是
38:13
breaking that down. And then I think with this iteration, right? It's going to happen a lot faster
把这一点拆解开来。然后我觉得，随着这一轮 iteration，对吧？这会快得多
38:19
and I think people will just naturally want to work together because I think I was at VM Explorer
而且我觉得人们自然而然就会想一起合作，因为我觉得我当时在 VM Explorer
38:23
and I did a few talks and I started every talk with like, all right, let's, let's get us
我做了几场演讲，每场演讲开头我都说，好吧，让我们，让我们 get us
38:28
show hands in the room who has used one of these tools like in at work, like cursor, claw code,
在座用过这些工具的人举个手，比如在工作里用过，像 cursor、claw code 之类的，随便哪个都行。
38:33
whatever. And pretty much the entire room and all of them went up and I was like, well, that's a lot
结果基本上整个房间的人，所有人都举手了，我就想，好吧，这比我预期多多了。所以你看，就连 adoption rate，在，你知道，那些传统的，你知道，
38:37
more than I expected. So like the adoption rate even at, you know, the traditional like, you know,
platform 或 developer engineer，还有，你知道，或者只是 infrastructure admin，在一些这样的 personas 里，也已经深入到那个范围了。
38:41
platform or developer engineer and, you know, or just infrastructure admin at some of these
所以我觉得，整个 enterprise 接受这个真的很快，而且，你知道，协作越多，我觉得就越有帮助。
38:46
personas is as well into that, that scope. So the entire enterprise I think is picking up on this
所以，你知道，就多参加 office hours，你知道，主动联系，主动一点。
38:51
really fast and the, you know, the more collaboration, I think it just will help out. So, you know,
然后我猜，如果你在，如果，要，要，要，要留心，或者对变化保持开放。
38:57
just office hours, you know, reach out, be proactive. And I guess if you're on, if, be, be, be, be mindful
而且这只是总的来说，如果我之前看到的情况是这样，如果你……
39:05
or be open to change. And just this is just in general, if what I've seen before and if you're
或者要乐于接受变化。而这总的来说就是，如果我以前看到的情况是这样，以及如果你
39:11
resistant and resist the change and, you know, that's, that's going to kind of get you labeled and
抗拒，抵制变化，而且，你知道，那，那会让你有点被贴上标签，而且
39:16
that kind of put off the side. But if you're, you may have your own opinions, but if you
那会让你有点被晾到一边。但如果你，你可能有自己的看法，但如果你
39:20
present them in a way that kind of is, let's say, maybe open-minded or say, you have a better chance
把它们呈现出来的方式，可以说是，也许更开放一点，或者这么说，你就更有机会
39:26
at some of those, those key learnings that you've dealt with throughout your career to actually be
让你职业生涯中处理过的一些，那些关键经验真的能
39:30
applied in the new way as well. That makes sense. I'm curious. I want to, I want to circle back for
以新的方式也被应用。这说得通。我很好奇。我想，我想先绕回来
39:36
a moment on something you said, as we're talking about kind of spreading across, my mind goes to
回到你刚说的某件事上，当我们聊到有点像在扩散时，我脑子里想到的是
39:43
security, because I know in, you know, in my industry security is a, is a pretty big deal. We do
security，因为我知道，在，你知道，在我这个行业，security 是，是件挺大的事。我们会
39:50
have air gap things and stuff like that is what we do in defense. But as you're, as you're thinking
有 air gap 这类东西，这就是我们在国防领域做的事。但当你，当你正在思考
39:55
about that, you know, one of the topics that is a really common topic now is kind of going back
说到这个，你知道，现在有一个特别常见的话题，大概要回到几周前。
40:01
a few weeks. We get, we covered it in depth on a previous episode that the whole open AI,
我们，我们在之前的一集里深入聊过，就是整个 OpenAI、swarms 跑出来、Hugging Face 那一整件事。
40:07
the swarms got out, hugging face, that whole thing. I, I'll refer people back if they don't,
我，我会让不熟悉的人回去听我们最近那期，他们讲过那个。
40:13
if they're not familiar with it, to our episode recently. They covered that. And, but that is top
而且，但这件事现在是大家最关心的。
40:20
of mind. It's not only top of mind for security, but it's top of mind for legal. It's top of mind
不只是安全团队最关心，法务也最关心，传播团队也最关心，因为他们在想我们该怎么办？
40:27
for comms, as they're thinking about what do we do? It's something like this where to happen in
如果类似这种事发生在我们组织里，不管那个组织是什么。
40:32
our org, whatever that org is. And can you, can you talk a little bit about, you know, as you've,
还有，你能不能稍微聊一下，你知道，随着你把企业结构更新到能应对 agents，而且，你知道，
40:39
as you've kind of brought the enterprise structure up to date to deal with agents and, you know,
当你某种程度上把企业结构更新到最新，以应对 agents，而且，你知道，
40:45
it's, we're beyond just the, the traditional software and app world. And we're, we're,
这，我们已经不只是那个，那个传统的软件和 app 世界了。
40:49
we're taking some of the same lessons and reapplying them in a new context here. Can you talk a
而且我们，我们，我们是在把一些同样的经验拿过来，在这个新情境里重新应用。
40:54
little bit about how that helps on the security side? If, if, for the folks out there that are
你能稍微聊聊这在安全方面是怎么帮上忙的吗？
41:00
really, you know, that's their job is to keep things from really going awry. Can you talk for a
如果，如果对外面那些人来说，他们真的是，你知道，他们的工作就是防止事情真的出岔子。
41:05
moment about keeping things in the box, if you will? Yes. Sure. And then I guess I'll, I'll put a
你能不能花一点时间聊聊怎么把东西控制在盒子里，可以这么说？
41:10
disclaimer. These are all could be suggestions to help out with. If you, if you really dig into some
可以。
41:15
of the, the, the hugging face stuff, it's a little like scary. So I'm not to say that we're like
当然。
41:20
better than all of the I labs or anything like that. So, um, but if you, if you look at some of the,
然后我想我会，我会先加个免责声明。
41:25
the things where how that kind of got started is where they, you know, they started getting
这些事情，就是那种怎么开始的，就是他们，你知道，他们开始从
41:28
out of the sandbox, right? And then they, they found things that were like, they could get access
sandbox 里出来，对吧？然后他们，他们发现了一些东西，就像，他们能接触到
41:33
to where, you know, having monitoring control around what those apps or agents can do is, is, you
那种地方，你知道，对那些 apps 或 agents 能做什么进行 monitoring control，是，是，你
41:39
know, something that's been it, been in mind for all enterprise apps, right? So like having,
知道，这一直是，一直是所有 enterprise apps 都在考虑的事情，对吧？所以就像有，
41:44
you know, kind of guard rails, even like network zones. So like, hey, this is a highly regulated
你知道，某种 guard rails，甚至像 network zones。所以就像，嘿，这是一个高度受监管的
41:47
zone. It's, it's super locked down. Like, even just like traditional firewall would have
zone。它，它被超级锁死了。就像，甚至只是传统的 firewall 本来就会
41:51
potentially blocked a lot of this stuff, right? So it's kind of key. Like, a lot of the core
可能已经挡掉很多这类东西了，对吧？所以这还挺关键的。就像，很多核心的
41:57
enterprise learning to kind of maybe take that for like, you know, the, the old team versus new
enterprise 正在学着，也许把这一点用起来，就像，你知道，老团队对新团队
42:00
team. It's like, well, guys, you could have just, you know, I don't know, had some basic controls.
团队。就像，呃，伙计们，你们本来可以，你知道，我也不知道，搞点基本的控制就好了。
42:04
And like, I think they even talked about like, their monitoring wasn't working. And they weren't
而且，我觉得他们甚至还聊过，他们的 monitoring 根本不起作用。而且他们也没有
42:08
really monitoring it and like, you have all these agent forms. And then of course, and then I got
真正在 monitoring 它，而且，你有一堆这样的 agent forms。然后当然，然后我就
42:12
to the Artifactory, which then found the weak link and the Artifactory had the internet access
弄到了 Artifactory，然后它找到了薄弱环节，而 Artifactory 有 internet access，
42:16
and somehow they magically, you know, zero data, the Artifactory and since multiple times. So not
然后不知怎么地，他们神奇地，你知道，把 Artifactory 的 data 清零了，而且之后还不止一次。所以不是
42:21
to say that we could prevent everything, but it's just, you take those, at least core enterprise
说我们能预防一切，但只是，你把这些，至少那些核心 enterprise
42:25
technologies that maybe would have helped out a little bit of just, you know, agent control,
技术用上，说不定能帮上一点忙，就只是，你知道，agent control，
42:30
like hardened sandbox type constructs, very locked down network controls. That would help
比如 hardened sandbox 类型的构造，非常锁死的 network controls。那会有帮助
42:37
a little bit potentially, you know, just, you know, monitoring what's in the, in the agent,
可能有一点点吧，你知道，就是，你知道，监控一下 agent 里面有什么，
42:43
make sure that the agents start, you know, somehow injecting tools that they're not supposed to have.
确保这些 agents 开始，你知道，以某种方式注入一些它们本来不该有的 tools。
42:48
I don't know, not sure if that was an issue or not. But, you know, when you talk about an application,
我不知道，也不确定那算不算个问题。但是，你知道，当你谈到一个 application 的时候，
42:52
the whole dependency stack is always up to, up to grab. So you want to make sure that the
整个 dependency stack 总是，嗯，总是说不准。所以你要确保，
42:56
dependency stack, the container, everything is in the whole OS and the whole platform is trying
dependency stack、container，整个 OS 和整个 platform 里的一切都在努力
43:00
to patched and reliable as possible. So it's a, I think the key thing is obviously, you know,
尽可能打上 patch、尽可能可靠。所以，我觉得关键点显然是，你知道，
43:05
we had the whole mind bending effect of that they somehow made this messaging board on Artifactory.
我们经历了那种让人脑子转不过来的效果：他们居然在 Artifactory 上搞出了这个留言板。
43:09
And then I got deleted and they recreated it. So there's that factor, which I think everybody's
然后它被删了，他们又把它重建了。所以有那个因素，我觉得每个人都...
43:13
a little like, and it goes into the whole AI is going to kill us thing that happened over the
有点像，然后就扯到那一整套 AI 会杀死我们的事情，那事发生在
43:16
weekend, right? Versus just do some basic security, basic stuff, right? Just to clarify that,
周末，对吧？而不是只做点基础安全、基础的东西，对吧？就是澄清一下，
43:23
that was a social media post, less people. Sorry. Sorry. Yeah, yeah, that's that's a quite yeah.
那只是个社交媒体帖子，人不多。抱歉。抱歉。对，对，那，那确实挺，对。
43:29
If you're not on AI Twitter or AIX, there's a lot of posts around, you know, somebody left
如果你不在 AI Twitter 或 AIX 上，你知道，到处都是帖子，有人离开了
43:35
anthropic and the much anthropic people started posting around that they're afraid that, you
anthropic，然后很多 anthropic 的人开始到处发帖说他们害怕，你知道，
43:40
know, there's like greater than 10% chance that AI is going to, you know, take out humanity over
就像有超过 10% 的概率，AI 会，你知道，把人类干掉，在
43:44
the weekend, slash Friday, and then that, that started up a lot of conversation. Let's just say.
这个周末，或者说周五，然后那，那引发了很多讨论。就这么说吧。
43:49
So that's what I was referring to. Yeah. Yeah. Okay. Um, I guess that, that's useful. I
所以那就是我指的事。对。对。好。嗯，我想那，那还挺有用的。我
43:56
appreciate that. I, we like to finish off by kind of giving you kind of a free form question.
很感谢。我，我们喜欢最后给你一个，算是自由发挥的问题。
44:04
And that is kind of toward the future, as you are, are doing your job and, you know,
这算是关于未来的，因为你在做你的工作，而且，你知道，
44:10
running your podcast and talking to people and, and kind of building up, you know, your
做你的 podcast，跟人聊天，然后，然后慢慢建立起，你知道，你的
44:17
perspective on where things are going with this. Um, we ask, we like to ask guests to, to go
对这件事会往哪儿走的看法。嗯，我们问，我们喜欢请嘉宾，就是
44:22
ahead and get crazy and prognosticate just a bit. And um, we don't hold you to it. But like,
放开点、大胆预测一下。然后嗯，我们不会拿这个来要求你。但就是，
44:30
really the way we do this is when you're, when you're kind of just letting your mind wander
其实我们做这个的方式是，当你，当你就是任由思绪游走的时候，
44:36
and you're thinking about these things in the back of your mind, you go into bed at night,
你在脑子里不自觉地琢磨这些事，晚上上床的时候，
44:40
having a glass of wine, whatever it is you do to chill out. What do you think things are going?
喝杯葡萄酒，或者不管你怎么放松。你觉得事情会怎么发展？
44:44
Like what, what, and you can choose the time frame that you like. But I really like to think like,
就像什么，什么，而且你可以选你喜欢的时间范围。但我真的很喜欢这样想，
44:50
what do you think is coming? What's next? Um, why, you know, what, what should, should orgs be
你觉得接下来会发生什么？下一步是什么？嗯，为什么，你知道，什么，什么应该，组织应该
44:57
having a thought toward for the future, even if it's not, even if we're not there yet today?
为未来做点思考，即使现在还不是，即使我们今天还没到那一步？
45:01
Um, any, any thoughts on what that future looks like as we close out?
嗯，在我们收尾的时候，对那个未来会是什么样，有没有任何，任何想法？
45:06
Oh, oh boy, it could be, it could be many things. But I, I think from a enterprise perspective,
哦，哦天哪，可能会是，可能会是很多事情。但我，我觉得从 enterprise 的角度来看，
45:11
I think we'll see a lot of like team change and I did roll chain and not even just an enterprise.
我觉得我们会看到很多像团队变化，还有我做了 roll chain，而且甚至不只是 enterprise。
45:16
Like the whole, like, from the business side, things are going to change a lot, right? Because
就像整个，就像，从业务侧来说，事情会有很大变化，对吧？因为
45:20
you start to see, and I was going through old episodes of my podcast this year and just some
你开始看到，而且我今年在翻我自己 podcast 的往期节目，然后只是有些
45:25
of the old, just the start of this year. And there's just now, I was like, wow, that's a lot has
旧的那套，只是今年刚开始。而现在，我就想，哇，变化真大
45:30
changed. And like, can't imagine, like, can you, like, you know, imagine a couple of years from now,
了。而且，就像，没法想象，就，你能不能，就，你知道，想象一下几年后，
45:35
how much change, the rate of changes now. It's accelerated. Definitely. So it's, it's hard to
会有多少变化，现在变化的速度。它加速了。绝对是。所以，这，这很难
45:40
imagine that. But I definitely think from a business perspective, we're going to see like a radical
想象。但我绝对觉得，从商业角度看，我们会看到一种彻底的
45:44
change in like how, how, you know, development and, you know, product management and everything
变化，比如，怎么，怎么，你知道，开发还有，你知道，产品管理以及所有一切
45:49
changes or everything kind of, you know, there should be like new roles get created, you know,
都会变，或者所有东西都，你知道，应该会有新的岗位被创造出来，你知道，
45:54
orgs are going to change, I think, and I think, I think for the good to you, because I think like,
组织会变，我觉得，而且我觉得，我觉得这对你是好事，因为我觉得，就像，
45:58
there's always a fear of a job loss. But I find that I've, and I've heard this echoed across many
总会有人害怕失业。但我发现，我，而且我听到很多地方都有这种共鸣
46:04
personal experiences, I am like, working harder than I ever have, because I have all these agents,
就个人经历来说，我现在就是，比以往任何时候都更努力，因为我手上有这么多 agents，
46:08
and they like, they need stuff from me. And I'm constantly like, and I'm doing building things that I
而且它们吧，它们需要从我这儿拿东西。
46:12
always wanted to build, but just never had the time to. So I feel like we're all going to be
而且我经常就是，我在做那些我一直想做、但就是一直没时间做的东西。
46:17
busy. And it maybe, maybe we need to, you know, not get too crazy with the agents, but it's hard for
所以我觉得我们大家都会很忙。
46:23
me, because it's fascinating. And it's really enjoyable. So doing a lot of things with automation,
而且也许，也许我们得，你知道，别在 agents 上搞得太疯，但对我来说很难，因为这太迷人了。
46:29
I can only imagine what these agents are going to be like in the, you know, the next six months in terms
而且这真的很享受。
46:33
of, you know, what they can do for you. I think we'll see a lot with voice come, come into play and
所以，用 automation 做很多事情的时候，我只能想象，这些 agents 在，你知道，接下来六个月里会是什么样，就它们能为你做什么来说，你知道。
46:37
maybe even a new style of interaction with the computer as well. So that will be interesting as well.
我觉得我们会看到很多跟 voice 相关的东西开始发挥作用，也许甚至会出现一种全新的跟电脑互动的方式。
46:43
And then obviously I think, you know, did you see the posts like hugging face head little robot
然后很明显，我觉得，你知道，你看到那些帖子了吗，像 Hugging Face 那种小机器人
46:49
things on sale? Yes. Yeah. Yeah. So like a little little micro dot things. I did. I was like, I was like,
东西在打折？对。是啊。是啊。就是那种特别小的 micro dot 小玩意儿。我看了。我当时就想，我就想，
46:55
I took my wife, I was like, maybe we should get what she's like, she's like, you have enough toys.
我跟我老婆说，我说，也许咱们该买一个。她就说，她说，你玩具够多了。
46:58
You just don't. Well, I may have ordered a few for the family. And I'm okay. There you go. Awesome.
你就是不需要。嗯，我可能已经给家里订了几个。我觉得还行。这就对了。太棒了。
47:04
So like, like robotics and things, like, you know, I think the LLM is great and wonderful, but
所以，像是 robotics 这些东西，你知道，我觉得 LLM 很棒、很厉害，但是
47:09
there's so many more possibilities where would you go that, I guess, like for humanity, right?
还有那么多更多可能性，你会往哪个方向走呢，我猜，比如对人类来说，对吧？
47:14
As long as you know, we prevent the terminator scenario, it'll be great, right?
只要，你知道，我们能防止 Terminator 那种场景，就会很棒，对吧？
47:18
Right. Absolutely. Yeah. Great. I appreciate you painting that. That that helps a lot. Yeah,
对。绝对是。是啊。太好了。谢谢你描绘了那个画面。那那真的帮助很大。是啊，
47:24
the rate of change is just insane now. And I don't I agree with you. I think it's going to get faster
现在变化的速度简直太疯狂了。而且我不——我同意你。我觉得它会越来越快
47:29
and faster. But great way to end the show. A lot of fun. I always I always love to hear what people
越来越快。但这是结束节目的好方式。非常有趣。我一直，我一直很喜欢听人们
47:35
say at the with when we're asking about kind of where they think things are going. Thank you for
在最后我们问他们觉得事情会往哪里去的时候会怎么说。谢谢你
47:41
coming on the show today. It was a great conversation. I learned so much. I really appreciate that.
今天来上节目。这是一次很棒的对话。我学到了很多。我真的很感激。
47:47
And I hope people tune in to Bug Boundary Weekly to hear you talking more about these topics.
而且我希望大家收听 Bug Boundary Weekly，听你更多地聊这些话题。
47:53
And they should definitely go to the Midwest Summit and hear your talk. I was there last year. I
而且他们绝对应该去 Midwest Summit 听你的演讲。我去年去过那里。我
48:01
did a talk. I won't be there this year. So I'll have to see if I can get ahold of the video of it or
做了一场演讲。今年我不会去那里。所以我得看看能不能拿到它的视频或者
48:07
something. But anyway, thanks for coming on. I'm proud to play out today. Thanks for having me.
什么的。但不管怎样，谢谢你来上节目。今天我很自豪能播出这期节目。谢谢邀请我。
48:16
All right, that's our show for this week. If you haven't checked out our website,
好了，这就是我们本周的节目。如果你还没看过我们的网站，
48:20
head to practical ai.fm and be sure to connect with us on LinkedIn, X, or Blue Sky.
就去 practical ai.fm 看看，也一定要在 LinkedIn、X 或 Blue Sky 上和我们互动。
48:26
You'll see us posting insights related to the latest AI developments. And we would love for
你会在那里看到我们发布与最新 AI 发展相关的见解。我们也很希望
48:30
you to join the conversation thanks to our partner prediction guard for providing operational
你能加入这场对话。也感谢我们的合作伙伴 prediction guard 为节目提供运营
48:35
support for the show. Check them out at predictionguard.com. Also, thanks to break master cylinder for
支持。去 predictionguard.com 看看他们。另外，也感谢 break master cylinder
48:41
the beats and you for listening. That's all for now. But you'll hear from us again next week.
带来的 beats，也感谢你的收听。现在就先到这里。不过下周你会再次听到我们。