Every Harness Will Become A Claw — Sam Bhagwat, Mastra

频道: AI Engineer
视频: https://www.youtube.com/watch?v=8qWIPUia2O8
原文语言: en
统计: 共 39 轮

[0:01]

[music] I'll get us started. Um, so long day of talks and uh, how are you all feeling?

（音乐）那我们就开始吧。今天听了一整天的分享，大家感觉怎么样？

[0:20]

Cool. Yeah, good to see we still have some energy. You know, I know there's a lot of like evening events. Um, we've heard a lot about the present and I'm going to talk about the future. Um, so my talk is called every harness will become a claw. Um, here's a little bit about me. Um, I am the co-founder CEO of Mastra. We are a Typescript agent framework. Um, I am also the author of a book that you may have gotten a copy of

不错，看到大家还有精神我就放心了。我知道晚上还有不少活动。前面大家听了很多关于「现在」的内容，我今天想聊聊「未来」。我这场分享的题目是：每一个 harness 最终都会变成 Claw。先简单介绍一下我自己，我是 Mastra 的联合创始人兼 CEO，我们做的是一个 TypeScript 的 agent 框架。我还写了一本书，你们可能已经拿到过一本

[0:45]

either outside or at a previous event. Um, we have seen a lot of agents running in production um, over the last 18 months. And I'm not I'm saying that as kind of context for and stage setting for the thoughts and ideas that I'm about to share right now. Um and uh the thing that I'm going to say is is welcome to the harness era. What do I mean by the harness era? Um well u let's just talk about the types of harnesses

在会场外面，或者在之前的活动上领到的。过去 18 个月里，我们见过大量真正跑在生产环境里的 agent。我说这个，是想给接下来要分享的这些想法做个铺垫。我想说的第一句话是：欢迎来到 harness 时代。什么叫 harness 时代？我们先来看看现在都有哪些类型的 harness。

[1:15]

that we see right now. We see local harnesses. Um we use them every day daily driving our our coding, right? Um we see cloud harnesses. Um these are both products that we can purchase as well as if we work and some of these uh companies that have built their own internal coding agents that live on Slack. Um and then of course we have the your friendly local uh open source frameworks that have some of these primitives and give you the the tools

首先是本地 harness，我们每天写代码都在用，它已经是我们的日常主力。然后是云端 harness，既有可以直接买来用的产品，也有一些公司自己做的内部 coding agent，直接跑在 Slack 里。当然，还有你身边这些开源框架，它们把这些基础能力都提供好了，让你可以自己动手

[1:41]

that you need to build your own. Um and that's where we that's where we fit in. Um now let's talk about where we are sort of collectively as an industry um and how things have evolved over the last we'll say year to to 18 months. Um last year at at AI engineer we were talking a lot about agents. We're talking about the agent loop. We were talking about agents versus workflows.

搭出属于自己的那一套。我们做的就是这一块。接下来我们聊聊，作为一个行业，我们整体走到哪一步了，过去这一年到 18 个月又发生了哪些变化。去年在 AI Engineer 大会上，大家还在聊 agent，聊 agent loop，聊 agent 和 workflow 到底该怎么选。

[2:03]

Um so so I want to you know there's as we're thinking about um the agentic spectrum I often compare it to uh self-driving as a spectrum right there are different levels of self-driving autonomy whether that's like lane assist whether that's Tesla S FSD whether that's I I'm sitting in the back of my Whimo and there's nobody behind the steering wheel right um there are various aspects of the agentic spectrum between LLMs agents harnesses and claws

在思考 agentic 这条光谱的时候，我经常拿自动驾驶来类比。自动驾驶是分等级的：有车道保持辅助，有 Tesla 的 FSD，也有我坐在 Waymo 后排、方向盘后面根本没人的情况。同样，从 LLM 到 agent，再到 harness，再到 Claw，这条 agentic 光谱上也存在各种不同的层次，

[2:30]

and I'm going to talk about what we've seen and where we're going. What makes an agent different than an LM? Hopefully, we mostly know this, but just as a quick refresher, right? It's the agent loop, its tool calls, its memory, it's the ability to retry failed tasks, it's context engineering. Um, Dex is a close friend and an inspiration for this uh one of the inspirations for this talk. Um, and it's agent state, right?

我今天要讲的就是我们看到了什么，以及接下来会往哪走。agent 和 LLM 的区别到底在哪？相信大家基本都清楚，但还是快速过一遍：是 agent loop，是 tool call，是 memory，是任务失败之后能重试的能力，是 context engineering。Dex 是我的好朋友，这场分享的灵感有一部分就来自他。还有就是 agent state。

[2:54]

These are some things that like hey I'm running an agent in a loop and I can't just do this with a oneshot call uh to to an LLM right I didn't I've tried to make these qualities I don't not sure what the quality of taking actions is active or something so I just put action but um you know qualities here are starting to emerge when we move from an agent to a harness durability and doggedness um a friend of mine was referring to an agent that he was using

这些能力有个共同点：我是在一个循环里跑 agent，光靠一次性调用 LLM 是做不到的。我本来想把这些都归纳成「特质」，但「能采取行动」这个特质到底该叫什么我也没想好，所以就先写了 action。总之，当我们从 agent 走向 harness 的时候，新的特质开始冒出来了：durability（持久性）和 doggedness（那股不撞南墙不回头的劲儿）。我有个朋友在描述他用的一个 agent 时，

[3:22]

and he called it dogged which I really like and I'm taking that for this talk, right? So, durability just the sheer quality of like being able to run not for minutes but for hours or days. Um, you know, what what what encompasses this? Well, sometimes it's like, hey, I you know, I uh lost a connection in the middle of the turn and uh you know, but I persisted the stream and so now I can resume from the place where I started,

用了 dogged 这个词，我特别喜欢，就直接拿来用在这场分享里了。所谓 durability，就是它能连续跑的时长不是几分钟，而是几小时甚至几天。这里面具体包含什么？比如说，我在某一轮对话中间断线了，但因为整个流被持久化了下来，我可以从断掉的地方接着往下跑，

[3:45]

right? There's planning mode. We all see this in cloud code. Um parallel sub aents being able to fan out multiple tasks at the same time. Uh we have more affordances with a TUI and slash commands. We have skills. Um we don't have to define all our agents up front,

对吧。还有 planning mode，我们在 Claude Code 里都见过。还有并行的 subagent，可以同时把多个任务 fan out 出去。TUI 和 slash command 也给了我们更多的操作方式。还有 skills。我们不需要一开始就把所有 agent 都定义好，

[4:01]

but we can dynamically create them on the fly. This is, you know, very powerful. Um you can sp the the harness can spin up background bash tasks, right? Um it will autocompact when it runs out of the context window. Uh you know, these are all things we'll see when we use cloud code or codecs, right?

而是可以在运行过程中动态地把它们创建出来，这一点非常强大。harness 还能在后台拉起 bash 任务。context window 用满的时候，它会自动 compact。这些都是我们在用 Claude Code 或者 Codex 的时候会看到的东西。

[4:20]

Um it persists. it will persist threads, right? You can resume a thread once that's you've like disconnected from later. Um, you can cue, you can steer, you can interrupt. You're not just blocked waiting on the LM. Hey, I take a turn and then you take a turn. I'm playing playing Civilization here and I can't take a turn until all the other civilizations are playing. No, I'm playing Starcraft. I'm playing Age of Empires. And um, right. Uh, you know,

它还会做持久化，会把 thread 存下来。你断开连接之后，过一阵子还能把这个 thread 接着跑起来。你可以排队（queue），可以中途给它调整方向（steer），也可以直接打断它。你不用干等着 LLM，不是那种「我走一步、你走一步」的回合制。那就像玩《文明》，别的文明没走完我就动不了。不，现在玩的是《星际争霸》，是《帝国时代》，是实时的。

[4:42]

session long um tool approval. So it's not just like yeah I approved this specific tool call but yeah you can run all instances of rmrf for slash right that you see in the session even though the first one will probably wipe your machine. Um okay so there there's like you know there's a um there's a few steps here and I'm I'm I'm about halfway through these and then afterwards we're going to talk about what it means and

还有 session 级别的工具授权。不只是「我批准这一次 tool call」，而是「这个 session 里所有的 rm -rf / 你都可以执行」——虽然第一次执行基本就把你的机器清空了。好，这里其实分好几个阶段，我现在讲到大概一半，后面我们再聊这些到底意味着什么。

[5:09]

this is kind of a in between step. I think this is something we've seen over the last really 3 months. Um, and I think we're all still starting to grapple with what it means, which is this movement from a local harness to a cloud harness where the harness is always on. What do I mean by a harness that is always on? Well,

接下来这个算是一个中间阶段。这是过去三个月里我们真正看到的变化，我觉得大家现在也都还在消化它到底意味着什么——那就是从本地 harness 走向云端 harness，harness 变成了 always on、一直在线的。always on 的 harness 是什么意思呢？

[5:30]

you might be talking to it in Slack. Maybe you're talking to it in Slack along with your colleagues, right? Um maybe you're each giving it instructions and has to figure out how to parse that and use user metadata. Uh maybe you have a mobile app. I was just uh uh you know maybe you have a mobile app. Maybe it tunnels to your local um to to your local machine. Some of these uh some harness mobile apps do this. Um often

你可能是在 Slack 里跟它对话，甚至是你和同事一起在 Slack 里跟它对话。你们各自给它下指令，它得自己搞明白怎么解析这些指令、怎么用好用户的 metadata。也可能你用的是手机 App。这个 App 也许会打一条隧道连回你本地的机器——有些 harness 的手机 App 就是这么干的。很多时候你会问，

[5:57]

like cool. How is this running? Well, it's probably running in a cloud sandbox because it's maybe it's running locally in your machine. and you're tunneling into it, but maybe it's just running in a cloud in the cloud somewhere and it's got a bunch of sandboxes which enables more parallelism. You can get more um parallel sub aents beyond what you can do on your machine. This is always a trade-off and always something you get

这东西到底跑在哪？大概率是跑在云端的 sandbox 里。当然也可能是跑在你本地机器上、你通过隧道连进去；但更可能就是跑在云上某个地方，背后挂着一堆 sandbox，这样就能做更多并行。你能同时跑的 subagent 数量，会远远超过本地机器的上限。这永远是个权衡，也是分布式系统必然会带来的东西。

[6:16]

with distributed systems, right? You can do more in the cloud than you can do locally. You have more resources. It requires a different architecture. It's more powerful. Um and then lastly, you're not creating code, you know, on your just on your local machine or maybe even in a git work tree. um you're you're probably creating, you know, if you're writing code, you're probably creating a PR that that pushes right to

在云上你能做的事比本地多，资源更多，但架构完全不一样，同时能力也更强。最后一点，你产出的代码不再只落在本地机器上，甚至也不只是落在一个 git worktree 里。如果你在写代码，你多半是直接生成一个 PR，推到

[6:35]

to GitHub. Um so so you know, there's a shift, right, from from local harnesses to these always on kind of like cloud harnesses. We're still in the middle of this. You may have you may only be working with a local harness. You may have started to see cloud harnesses pop up in your your organization, right? You may be figuring out how to use them.

GitHub 上。所以这是一次明显的转变：从本地 harness，转向这种 always on 的云端 harness。我们现在还处在这个转变的中途。你可能现在只用本地 harness；也可能已经开始看到公司里冒出各种云端 harness，正在琢磨该怎么用。

[6:56]

Your teams may be figuring out how to use them. Um, and then I want to talk about what the harness to claw transition is, which is imbuing these agents, imbuing these harnesses with initiative and and learning, right? What is initiative? Well, um, if you've used,

你的团队可能也在摸索该怎么用。接下来我想讲的是，从 harness 到 Claw 的这一步转变到底是什么——就是给这些 agent、给这些 harness 注入主动性（initiative）和学习能力。什么叫主动性？举个例子，如果你用过

[7:13]

let's say, a a personal assistant uh, agent, right? And that agent texts you and says, "Hey, I saw an urgent email come in. Is that email actually urgent? Was someone like, you know, spamming you?" you know, but like like the agent is listening um to external feed services. It has a heartbeat which means it wakes up every you know defined amount of time and um and does something right. Uh again like channels some uh

比方说个人助理类的 agent，它会主动发消息给你：「我看到一封紧急邮件进来了，这封邮件真的紧急吗？还是有人在给你发垃圾邮件？」这说明这个 agent 一直在监听外部的信息源。它有 heartbeat（心跳），也就是每隔一段设定好的时间就醒过来一次，然后做点什么。再就是通道（channel），

[7:40]

you might be able to text it, WhatsApp it, telegram it where whatever you want. Um you you may persist the memory memory in a more accessible later place than just simple sort of like file storage, right? um you you might it might have a a Damon, it might have a a gateway uh for for sending in and receiving incoming outcoming requests. Uh it often will do continual learning, right? So this concept that you know the agent the

你可以给它发短信、发 WhatsApp、发 Telegram，随你喜欢。memory 的存储也可能不再只是简单的文件存储，而是放在一个更容易访问的地方。它可能还常驻一个 daemon，有一个 gateway 来收发进出的请求。它通常还会做持续学习（continual learning）——也就是说，这个 agent、这个

[8:08]

harness runs and then you know based on the traces that it generates it it sort of autoimproves itself and there's different ways of doing this. you see um skill automatic skill generation for example is a common one. Um it could modify the code driving this as well. Um we haven't figured out what the right way of doing it is yet. We're still exploring you know the industry is still exploring options. Um now the reason that and and maybe this is like

harness 跑完之后，会根据自己产生的 trace 来自我改进。实现方式有好几种，比如比较常见的是自动生成 skills；它也可以去改动驱动自己的那套代码。到底哪种做法才是对的，我们还没搞清楚，整个行业也都还在探索。至于说原因——这可能是

[8:37]

our unique vantage point here but you know for the last three months as a framework we've just seen this as a f as the future. And so we furiously looked at the the you know the features that you know openclaw have that Hermes agent have and say and and we we've said like look you know a lot of people a lot of folks want these features but they want them with power and control. They don't want to just put a you know a claw on a

我们比较独特的观察视角：过去三个月里，作为一个做框架的团队，我们很明确地认定这就是未来。所以我们疯狂研究 OpenClaw 有哪些功能、Hermes agent 有哪些功能，然后我们的结论是：很多人确实想要这些功能，但他们想要的是有足够能力和掌控力的版本。他们不想只是把一个 Claw 随便丢进

[9:01]

box right they want to have more. And so, you know, we we've been thinking about this because we we you know, my I'm not doing my job well if I'm not giving everybody the tools that they need to build agents, to build harnesses as with the maximum power, right? Um so, so hopefully like again we hopefully I've walked a little bit through the step transition with actions, durability, doggedness, always on initiative, learning. Again, I I

一个盒子里就完事，他们想要更多。所以我们一直在琢磨这件事——如果我没能给大家提供足够的工具，让大家去构建 agent、构建能力最强的 harness，那就是我没把自己的活儿干好。所以希望我刚才把这一步步的演进讲清楚了：action、durability、doggedness、always on、initiative、learning。还是那句话，

[9:30]

think I failed in like making them all the right tense phrase and making them all qualities, but I hope you get the idea here, right? Um, we're ascending on the agentic spectrum. Um, and what was a simple LLM 18 or 24 months ago is a lot more powerful. So, I've called this without sort of asking consent from Pete, but I've called this Steinberger's law, which is I I believe every harness will expand until it becomes a claw. and

我承认我没能把这些词都统一成同样的词性和形式，但我想你们能 get 到我的意思。我们正在 agentic 这条光谱上一路往上走。18 个月、24 个月前还只是一个简单 LLM 的东西，现在已经强大太多了。我给这件事起了个名字——没经过 Pete 本人同意——我把它叫做 Steinberger 定律：我相信每一个 harness 都会不断膨胀，直到它变成一个 Claw。

[9:56]

and and that's a little bit um technological, that's a little bit economic, that's a little bit psychological. So, let me walk you through the reasoning here. Um the first thing that I've observed um that we've all observed um as a is that harnesses tend to expand. And they expand because we want them to expand. Um we want to DM them in Slack. We want to text them and like start overnight uh tasks before bedtime. We want this dopamine casino

这里面有技术层面的原因，也有经济层面和心理层面的原因。我来把逻辑给大家捋一遍。第一个观察，其实也是大家共同的观察：harness 天然就会不断扩张。而它们之所以扩张，是因为我们希望它们扩张。我们想在 Slack 里给它发私信，想睡前发条消息让它通宵跑任务。我们想要那种「多巴胺赌场」的感觉——

[10:27]

that we get when we put in tokens and get out code, right? Um or or whatever other actions, you know, um agent agents are bigger than just coding agents, but um we want our own dopamine casino. And this this image is thanks to uh to Dex Horty. Um but I see something else in our future. um which is that and and it's something that like I don't think we we sort of talk about as much uh but after this after this phase where where we're sort

投进去 token，吐出来代码。当然也可以是别的行动，agent 远远不只是 coding agent 这一类。总之，我们都想拥有自己的多巴胺赌场。这张图要感谢 Dex Horthy。不过我还看到了未来的另一件事，一件我觉得大家聊得不太多的事：在我们把一切都做得越来越强大的这个阶段之后，

[11:02]

of making everything more and more powerful um there will be a shakeout um and and let me walk you through sort of uh through my reasoning here which is that in the 2010s we had these platforms we had Android, we had iOS. And all of a sudden, there were all these things we could do on our phones that we previously weren't able to do. We could get directions, we could hail rides, um,

会有一轮洗牌。我来讲讲我的推理过程。2010 年代，我们有了 Android、iOS 这些平台。一夜之间，手机上能做的事情一下子多了起来，很多都是以前做不到的。我们可以查路线导航，可以叫车，

[11:30]

we could send payments. Um, we could play music. Um, other ones emerged over the course of the decade. We could watch short form video. Um, we could [clears throat] watch long form documentaries. We could browse the internet. You know, some were kind of ported over from the desktop. We could browse the internet. Um, again, you know, some, you know, we could order food, right? Um, we could put book accommodations, but but if you look at

可以转账付款，可以听音乐。还有一些是在这十年里陆续出现的：可以刷短视频，可以看长片纪录片，可以上网浏览。有些能力是从桌面端搬过来的，比如上网。还有点外卖、订住宿这些。但如果你去看

[11:54]

most of these kinds of categories, and there are quite a few categories, there really only like one or two, you know, logos here that we use, you know, okay, how many maybe, you know, we use Uber and we we use Lyft, but like do anyone use another rideing app here, you know,

这里面绝大多数品类——而且品类还真不少——每个品类里我们真正在用的，其实就那么一两个 logo。比如打车，我们用 Uber，也用 Lyft，但在座各位还有谁在用第三个打车 App 吗？

[12:10]

like and and so when you talk to people that are smart about like consumer behavior, the the The reason they say that this is is because you only really have space in your brain for like a limited number of things. Like if if think about something like Thumbtac. So uh Thumbtac didn't really serve a very high economic value. Like Airbnb like we only use it Airbnb occasionally, but when we use it, we like really want it. We really need it.

你要是去跟那些真正懂消费者行为的人聊，他们给出的解释是：人脑子里能装下的东西就那么多。比如你想想 Thumbtack 这类产品，它带来的经济价值其实没那么高。再比如 Airbnb，我们只是偶尔用一次，但真要用的时候，是非用不可，是真的需要它。

[12:41]

You know, it's really valuable to us. Thumbtac like a little bit less so, right? Um and then it's also like not frequent, right? like maybe maybe like you know something like Door Dash or Uber people can use multiple times a day, right? So there's there's sort of like it either has to be very economically valuable or has to be very frequent. And if it's neither one of the two, um we just forget about it, right?

它对我们来说价值很高。Thumbtack 就差一点，对吧？而且它的使用频率也不高。像 DoorDash 或者 Uber，有些人一天能用好几次。所以说，一个产品要么经济价值特别高，要么使用频率特别高。这两条都占不上，我们就把它给忘了。

[13:03]

It's like that, you know, college friend that like we haven't really talked to in years. It's not cuz like they weren't important at one point in our life, but like there's just nothing that maybe they moved to a new city or we moved to a new city or our lives, our friend groups, our careers diverged and all of a sudden like, you know, maybe we're calling them once a once a year or once every other year or or whatever. And

就像大学时候那个朋友，已经好多年没联系了。倒不是说他们在你人生的某个阶段不重要，而是可能他们搬去了别的城市，或者我们搬走了，生活、朋友圈、职业路径全都各奔东西。突然之间，可能一年才联系一次，甚至两年一次，反正就是这样。

[13:25]

you know, there's just nothing that makes them pertinent and brings them up in our our brains. And so right now we're like really excited because there's all this energy and excitement and we're all excited to these harnesses that we're like, you know, putting in tokens and getting out like useful things that we all love. Um and and I think that in the notsodistant future there will be this very real shakeout and and these categories will

就是没有什么事情能让他们跟你还有关联，能让他们在你脑子里冒出来。所以现在大家都特别兴奋，到处都是这股能量和热情，我们对这些 harness 都很上头——往里扔 token，出来的全是我们特别喜欢、也真正有用的东西。但我觉得在不太遥远的将来，会有一次非常真实的洗牌，这些品类会慢慢地

[13:47]

kind of emerge and we'll realize that we only have space in our lives for so many of these claws. They're very powerful. We we love them very much. Um and so I would I would think about um what what what does that mean for you? So the the first thing is um the first thing is don't get this is the reason that like events are are are important that like staying up with like if the rate of change increases 3 to 4x that means you know we

浮出水面，然后我们会发现，我们的生活里其实只装得下这么几个 Claw。它们非常强大，我们也非常喜欢它们。所以我会去想：这对你来说意味着什么？第一件事就是，别掉队。这也正是这类大会重要的原因——你得跟上节奏。如果变化的速度快了三到四倍，那就意味着我们

[14:19]

need to figure out what's going on even more frequently. That's why we're all here. Um, but if you're building an agent, make sure that it has the capabilities that your users need because if it doesn't and if there's newer things that come out, like they may just, you know, pick pick something that's more powerful because that that's happening very quickly. Um, and then keep in mind that if you if you if you aren't if you're the thing you're

得更频繁地去搞清楚到底发生了什么。这也是我们今天聚在这儿的原因。另外，如果你在做 agent，一定要确保它具备用户真正需要的能力。因为如果它不具备，等更新的东西出来，用户可能转头就去选那个更强的了——这种事现在发生得非常快。还有一点也要记住，如果你手上做的这个东西

[14:40]

working on, um, if even if you climb up to the top of the hill, keep in mind there's going to be another wave of this sort of these like this this shakeout coming. and you know probably sometime in the later 2020s. Uh so um that I'm Sam um I'm the uh I'm the co-founder of Monsterra the the TypeScript agent framework. I'm the author of Principles of Building AI agents. Hopefully you can get a copy of the book outside or I've

就算你已经爬到了山顶，也别忘了，下一波这样的洗牌还会到来，大概就在 2020 年代的后半段。好，那么，我是 Sam，我是 Mastra 的联合创始人，Mastra 是一个 TypeScript 的 agent 框架。我也是《Principles of Building AI Agents》这本书的作者，希望大家能在场外领到一本，我手上这儿也

[15:08]

got a few here. Um please stop by, say hi. Um it's great to see all of you. Thank you all for coming out. It's a real pleasure. Um enjoy the rest of the conference.

还带了几本。欢迎过来找我打个招呼。很高兴见到大家，谢谢各位今天来到现场，真的非常荣幸。祝大家享受接下来的会议。

[15:32]

[music]

[音乐]