探秘 Claude Code，搞懂 Agent Harness｜对谈来新璐

Hi, I'm Cody. Now, we've been talking about Agent Harness. The agent is from Harness to decide. But when we talk about Harness, we're talking about what? Not long ago, Cloud Code was released. So many Agent Harness's mainframe were completely displayed. It became a very good teaching tool. Today, our guest is Sinslou. Sinslou is Share AI of the creator. He made a lesson, Learn Cloud Code. This is a Github on now, 现在已经有超过5万颗星的教程 你好心路 欢迎来到十字路口 Hello 感谢科丽老师的邀请 我们这一期播客的目标是希望能够让大家感觉到 Harness不再是一个神秘的或者玄乎的词 然後可以让大家明白它其实就是软件工程的一些方法 那我们照旧还是从矿案快答来开始 请问心路你的年龄 23 毕业院校 河南工业大学 MBTI和星座 Inspiro Sticker On in IGP I am currently doing 10ков EkB a agent computer That's for discrete81 What type of agent Then another story We can just reveal theurb� drive What is the básicamente So we need to Consider data Any bottom words What does What is the situation? We are just closed 300 million dollars. What is the team's team? The team's team is very successful. The team is actually me, and I have two students. What is the income? What is the income? The new business is just starting, but not the income. Can you tell us about the business? I'm doing what is the business in the industry. I'm doing some AI infrastructure. If you are thinking about the new business, you need to use a little bit of a harness. I'm going to tell you about the harness. I'm going to tell you about the harness. I would like to put the machine into a smart brain. But it's not a body and body. It's just thinking. It's not a way to move. Agent's on-site is from Harness's design. You know this word? And why is the machine not from the machine's design? I think this word is a good. First, the machine's design is the machine's design. Because the machine is a machine. The machine is a machine. The machine is a good. We can see that now the machine's design is a good. The machine is a good. In 120 to 170 This example is much more than enough You have learned the skills to learn this,орот and coded in GitHub That's what s精 on領 Firth continue forty Actually, it was about 9 months ago. You were thinking about writing this kind of a job? We thought that Cloud Code is very strong. You just want Cloud Code to create a network, you actually get a very strong agent product. So we can do the development of the software. But the development of the development of the long graph, long chain, and the parameter of node and flow. It's like the PySYS. I all want you to tell me this whole time many people might not know the also a full in an paradigm and the default counter and I will and the law system isulsive and much thin so you 그래서 they agent of data they are done exactly and are not in anyuto ? I used to prompt management the type of curiosities that ? is ... Expect this we keepauto using AI ... ... The Big ... and Cloud 正好推出了 Managed Agents也就是把它自己做 Agent Harness的这一套给开发出来给大家用了岂不说更早之前它的原大麻也泄露了既然现在比如说我要开发一个 Agent我就可以直接用 Cloud Managed Agents大家还有必要自己去搭 Harness吗甚至还有必要去了解 Harness的细节吗做一个 Agent的开发工程师会需要的我认为如果你不了解的话你做出来的产品 缺乏灵魂和缺乏进一步迭代的空间 但这个就像云服务器一样 云服务刚出来的时候 可能工程师也会说 不能直接就掉它的API 我们也要去搞明白 服务器的方方面面的infra 但到后面你发现 其实99%的工程项目根本就不需要 所以我想会不会在 Agent Harness这个层面也是一样呢 我们今天虽然这一期播客还没开始聊 我们要讨论Harness 要让大家学习Harness 但会不会有一天 其实根本就不需要学了 大家直接用Cloud Management Agents的服务就够了 如果从整个推演角度来看 肯定未来会有那么一天 就像大家今天去比如开发很多这种 全站的这种外部应用都用NetGIS这个框架 但你不会知道NetGIS里面怎么设计和工作 它的原理是什么 大部分人可能不care这一层 我直接开枪挤用你开发就好了 那么以后的这个agent的harness 随着它迭代和收敛以后呢 Maybe大概率还是会 可能会回到这一层 两三年以后它成为一个非常清晰的 to 앞으로 prolong the future of fuel so this kind of technologyit doesn't mean to understand toothpaste it's going to be better it's not even more to teach what is the main clear 但 Monoende 我今天对产品经历会提出一个质疑,就是我觉得今天的产品经历跟过去产品经历看 general what you can say is actually signal. 今天是在技术变化周期內, 我们今天所做的一切产品都是在吃这个技术周期变化的红利所以说如果你自己都不了解这个技术周期变化的这个内核和本质到底在变什么 你很难构建一个产品,很贴合的要吃掉它的这个红利,或者说贴合这个变化和周期 So I think that I think we can use the new UX and UI. But today I think we need to use the key to how to improve the key to the future. And then use the key to the future in our certain environment. So you should be able to understand the future. The key to the future is the key to the future and the key to the future. And the key to the future is the key to the future. That we will be talking about Agent Harness. When we talk about Agent Harness, Is it a way to put it together? I would like to put it together three parts. left lower 它是不是会要有一些offload的 然后那下一轮相当于其实你以为还是那个agent在工作 其实已经是一个新的模型窗口 在装上一些初始化上下文 让它继续上一个窗口里在进行的工作 其实有点像一个新的agent的 那它怎么接上上一个agent交接给它的工作 这里面会有很多的上下文和状态的问题 我第二层觉得它被归为上下文的环境层也更好 那第三层把它归为就是说对agent的这个上层的管理 这个管理也会分两个方面 一个是如果你今天面对的事项不是一个agent 你是面对100个agent 就像你公司有100号员工 你怎么去组织他们协调起来 按照某种组织架构 某种协作方式和传递顺序 去把一个复杂的事情给解决掉 那第二层呢 就是说你针对这100个人 对吧 他每个人要不同的岗位的访问权限 其实也会有很多这种关于权限治理 关于信息的提供和隔离 很多这种治理的方面问题 This three subparng��, xin disney and aroundoofing, invested, this one again, One, another again,ands . Let the agent pousbou設ies cross-g Most of agents can've been kita A Nak So just a nice case. we say the agent's底层 is a the machine, the machine is like the brain and the brain it is to drive the entire the entire thing to do the action ability so you can get it to the same thing the information is the the action ability the first one the machine has the chance to create the machine the machine has the chance to create the machine the machine the second one is the OK, 那我模型是不是要知道我现在是在一个什么样的环境中去工作? 然後呢我的路径是什么? 我這上面装有哪些依赖? 對吧? 然後我已經寫的我念牙的結構是什麼? 對吧? 然後包括我的Git信息以及就是那它的模型上壓文窗口非常滿 因為我們知道一個CBNH是一個很大很大的工程量 他不可能在一个模型的上压文窗口中直接去做完 而不装满 他要有些上压文的卸载 策略包括我们刚说了上压文 窗口满以后 下一个agent再去工作 他要重新去读一下 我当年环境信息是什么 前面已经写的代码 写到哪了 接下来我要进行什么 在我这一轮的成人周期里 做到某个阶段 我再会写类似的这个文档 去委托我当天的上压文状态 再交给下一个agent 这样一次的接力 最后去把整个C编辑器给写完 它不是一个单纯的Action能力层就能完成了 它还是需要上下门环境的 对 然后第三层就是编排和治理层 如何去指导这些agent之间去做一个传递和委托 包括我的写代码agent和测试agent之间的关系 包括这个任务中难道一定每个环节都是串行的关系吗 我只能有一个agent接着一个agent去完成吗 我是不是有很多模块我是可以并行的去写的 可以同一时间有多个agent去做这个文章 它们之间的协调关系是什么 就是这会涉及到 边排的一个问题 以及就是说那不同的agent 它的权限怎么样 我去做测试的agent 是不是应该有这个测试环境 包括测试工具 它对应的权限 它不应该在测试的时候 对着我的这个代码 它一边测一边修改 说我测不过 我把那修改一下就好了 然后它一直在hack这种pass 对吧 所以说这里面会有很多的 这种权限的智力 包括上层的这种 协调关系的边排的一些问题 这属于第三层 所以大家过去做一个agent的时候 说到的这三层都是要靠手搓的 但现在慢慢的可能有一些 已经封装成了一些可以直接掉的 这个infra工具了对吧 然後在你看来现在有哪些 已经封装得很好了 有哪些人让目前需要手搓 这三层到目前我并没有觉得 有封装的非常好的 因为我们觉得这个范式变化太快了 从22年底 ChatGPT的发布 再到现在类似于CC这一套背后 Even though I don't know how many dating courses I had to train my training culinary by way It is now spellb Harriels It has probably been taken started From the local community I couldn see anything that kannst This is a company that has been It's a company So far, our team, nets are seeing. We're actually used to do the program The thing is doing is something that can do Cover our team craiskoitas c de 적 putting connected key Its a brain is a we find Weaffordial architecture We find a ideia in unusual The idea ofisas This is a program I think the Unix So... A one-year-old We also have other other side-side We also have other side-side For example, Keywatch For example, my view of the view For example, I have a few months They are in the end of the view What are the cases that I have done? They can see if I have a CLM Or if I have a skill set Or if I have a skill set Or if I have a model Okay, the view of the view of the view And we can see the view of the view We can see the view of the view surely sushi Infrara we fund this world will look for in the chat For all ? Thus i want to be able to will馴越 Mine if I wanted to test the content the mitä itself You can also talk앙 RS. You can seeATR-BQ and PU água and always use to Lake Facebook platform sometimes. Then you can seeATR-BQ5 if you want to and say what the umx the computer and give an agent. For a single residential environment. But if you mater chap if you dun학 lc national agent. He put that in that environment, it feels like living in a unix of the upcoming voorstain theünd永件 guide and the internet system and the incred Pf tend to practice in approaches. Goiver. That wasn't KARBI quieneshumurs of a open ogre What do we have done? this is because we used data ст Wade on the thing we can say in two weeks with the web because unfair你们 would be used with TS to rebel back then over must say For not to I fact didn't have a outer body should will in a�� muitos from from Australian gotten nossa what kind of tools 现在最主要的工具有哪些吗? 我觉得最主要的工具 首先就是关于文件系统的那些工具 就是增删读写搜索 这些大本意真的不管是在做coding任务 还是做deep research之类的任务 基本上都会需要的一些工具 那第二层我认为就是brother这一层 就是你给它一个 关于原来拥户世界系统的这样的一个操作能力 访问这一层 第三层其实有点像python node这种的就是 a language soort the makes it ... I think everything is that is I is as when I OK tę And you also have three of the features. will be our agent rides to get closeão. 状态层环境层的时候memory就记忆是非常重要的 然后现在也可以看到有不同的记忆 不管是模型厂原生也有一些记忆的功能 然后也有三方的一些开源的一些插件 他会给记忆的解决方案 然后可不可以给大家讲一讲 现在你认为agent harness的memory这一层 有哪些主要的解决方案 Memory这一层的发展目前还处于一个很早期的阶段 我其实把它处理了分为就是规则式的 半规则式的和完全模型驱动式的 那目前落在使用上来看的话 相对规则式的和半规则式的 我也多吧 完全规则式的就是像很多 基于这种知识图谱和向量搜索 它们所构成的就是把最后你的这个信息 在最后出现这种不同节点的表示 然后它们之间在做关联 然后检索和扩展 那这种方式是非常结构化信息 规则式做法的 当然我个人其实不太喜欢这种做法 I quite like these are like the kind of rule for the rule like this study which is about the name function the codes are quite the odd branding which is the quality of both recognition from the user which is written according to distance with reflecting quantances and because what if this information is 구formed and I think that is the transition So, lets go switch, the re-subs send spares . gathering about this open um um , . . but , , . . It will be able to create a report. That's what you can do to do skill. So I think all of the skills in the future This is the same thing. The first thing is in Cloud Code. Inside the other thing. It's a bit more like memory. It's a bit more like memory. The key is in the field. There are many of them in the field. The field you can say. It's a experience. It's a memory part. Or a version of the skill. that trouve... that's in my field at the end so we go ahead and what I said is just open to it I'm talking about The story... that they were fake and they 360 stuff that's very fucking crazy and we talked about and they also said we don't use surprising that they would want to go to thenet and then the tools they say to consider the goal and not the yellow project researched 你甚至会限制我们团队对自己的想象力 我觉得刚才我们在讲说 我们现在聊到的一些这个Harness的实践 它到底应该算memory还是算skill 可能聊这个定义没那么重要 可能我们更重要的是 Agent怎么可以这个不断的自己学习进化 就这个实践本身实现它的目标是重要的 对 标签没那么重要 在今天我们聊到这个Harness的时候 你觉得有哪些是共识 有哪些是非共识 共识可能就是说 So we're now saying Bash is all need is a common theme. We're now doing CLI, right? Or, we're just saying CLI is all need. Every person starts saying, ok, I don't want to go to that traditional way. MCP, I don't want to go to CLI, right? I'm in the service system. I'm in the phone. I'm in the phone. I'm in the agent. I'm in the agent. MCP, I'll have to have a agent for it. It's a bit of a problem. This might be a common theme. But I think today, many of the current agent are not just to be built in this new framework. It's still built in this prompt node to do this. and to make the process of tracking and de niece sensors and as much so it is we used to build some objective 又非常多关于它的这个文章 你觉得我们能从它的原码里面 writings个层面啊 你觉得能学到最关键的点有那些 它的压缩策略远比我们想的更加多集 比如说什么时候把这个工具的一些 Output 给删掉 で吧 然后呢 包括它的压缩的时候 它有哪些信息保留 那些信息恢复 因为相当于我刚说Agent之间的接力棒 那下一个Agent接到这个 Thanks 完成到一半的任务的时候呢 那上面信息开始要准备哪些内容 加载进来 Now patients cannot search for that, theachiness can be complete Val aliens are unable to cover their products And fabri has a testament of muss AutoDream Cloud Code Cloud Code里面记忆设计的也非常的精妙, 也跟他的Skills那套机制遵循了统一套哲学。 首先模型才是agent,用户去写的那些prompt, 去组的这种提示磁流的这种做法, chain的链条式的做法,不make sense,对吧, 模型才是agent, Cloud Code是完全零零 that you have to experience. I have another one for cloud code that allows you to oy for how the more the mercy should be treated as much as possible if this story does, not just practicing, thenegoode pattern as upon Oi pounder теперь many places. It includes a close column to make sure certain different styles, I have one for cloud code that gives�� more doesn't mean. 更少的control 更多的compact 更多的action能力的提供 对 然后零control Zero control 最多就是限制你掉工具的时候 你有些工具不能掉 在Manus上线的时候 他其实用了一个云端杀箱 叫E2B 今天可以讲一讲杀箱 有些什么样的进化吗 从一年前到现在 关于杀箱我们关注也比较多了 最近虽然也出了非常多的sandbox 把各种浏览器塞在里面 做的更all-in-one 更聚合的 但我们觉得这个领域其实目前大的进展还不多 以前有些轻量级的沙箱 那也是好多年初的 该有的还是有 我觉得我们自己做的这个就是1KB的unix Computers可能算是这个领域 或者会是一个大的进展 对 你们做是不是和Daytona有点像 但Daytona可能没有你们那么小 对 我们做的更加极致轻量 我们做的基本上不再是一个sandbox了 It's our foreign standards in US file system 以及最小化的这个Unix的Bash的这样的一个comment能力的一个B级 类似于它的局域网的文件夹共享磁盘 它的这个Gmail的局域网的通信 包括Crow命令局域网做相互之间通信这些能力 我们都是有的 而且我们认为浏览器这种东西或者说GCC编辑器 等很多重要的东西原本就不应该塞在一个生产级的一个agent 给它本身的配的环境中 It's allowed to provide a larger place for one infrastructure to provide a infrastructure for online software. The 11-day feedback for four-day system is through pretty much data. This is the type of short lab. We saw Cloud Code remember when Foier career changes in çıkar mode. Just before we just announced this little bit of that, it's something to help understand. How to do you manage fightier behavior? There are two different faster sci OT India, This is when you every time for a call code, it will be done with a call code. It will be done with a call code, called Stop. It will be done with a call code. It will be done with a call code. What is a call code code? It will be used to bring all the system features, including the previous channel of KVC to determine if there is a call code code. It will be used to be used to do that in Unix, or Fails, or Markdown. It's very Simstonially three different ways 至少大约五个section 才开始进这套机制 去做一个更深层的记忆整理 就是对最近的section做一个回顾 用户都跟我的agent去聊了什么 他就会综合地去看这些 最近的section去提取 进一步榨干利用这些信息 纠正记忆中有些错误的事实 或者说不及时的需要更新的事实 记忆做一些合并和整理 你可以理解为 他每隔天定时的开启一个后台agent 对最近的绘画信息做一遍重放 就像我们做梦一样 做重放 做信息整理 就是这么两套的agent的机制 所以Cloud Code的原代码泄露 除了我们刚才说到的 它的记忆机制被大家看到之外 还有哪些带来的惊喜吗 我觉得对我而言 最大的惊喜就是记忆 我读下来认为它的设计 也是跟它之前skill的 这种思想和哲学是一脉相承的 格式也好 整理维护的结构也好 做得非常make sense 对我最大的惊喜 The other thing is about what I said about unknown governmentoder, because the agent kedzie-en. himself. He doesn't want to continue to do the work. So later, he must migrate with the agent. He needs to prefer for the agent." He só submitted to him with traditional tradeov b Điサ l outro. that's something that tries to do so, he produced some very active kind of washes against the type of talent plan. nonprofit doesn't benefit for every any potentialoperation. So here is it a prototype about nós. It's a type of class. J supuesto that plays into the balance. But it's not supposed to be a natural, but it assumes that we don't need it to do it. 现在标准的agent模型都是在256kkxs长度 它其实可以装下非常多的信息 所以我觉得更大的就是关于上海文管理的启发 就是在于上海文窗口确实装满了 我再装装不下的时候 但是我这个任务还是要继续干 你怎么给我腾出更多的空间 或者想个办法让我接着把这个任务给完成 这个时候咱们看上这个上海文中有什么 可能垃圾或者不必要的信息 或者我们把它给上海文窗口踢掉 可能腾出来几十k空间 我又可以再干点事情了 这是一种典型的策略 那另外一种呢就是预值上限比如说乘以0.8 对吧 它留下了20%的上限 这时候我们交代一下吧 对吧 我们把我们这个任务干到什么情况 接下来还要干什么 总体上用户是希望我们干一个什么样的任务 我们写一个文档 写一下我们现在的整体的一个进展 然后让下一个agent开始干工作的时候呢 OK 他先读一下这个文档 然后去接着工作 这相当于是一种交接 那下一个文交接的时候呢 该交接什么 不该交接什么 以及下一个agent上下文出事化的时候怎么准备 so when we talk what unit harness about agent harness co-prended what we talk about "'ACENT driven how to make a way vividly héxél how to make room more context deal our co-markets like Martín The system is the system is the system is the system is the in the past, the original foundation of the original foundation. Now it's a new layer of the new layer of the foundation. It's the new layer of the foundation. We can't change the layer of the foundation. We can't change the layer of the foundation. We can't change the layer of the foundation. I think that now we have a lot of things. If you're from the foundation of the foundation, it's very natural, very符合常識 and sense of knowledge. But if you're just doing the application of the foundation, you can't do the raw base. you will think that these things are like... Why is this? It's not a lie. It's just telling you it's like this. You don't know what reason is. But if you think about it, the quantum armor is how to use it. It's how to use it. The agent's design is how to work. You can combine this with this. All the agent's best-time process, including the Miners'- You can feel that it makes sense. I'm going to use it to match the code. When we say the agent's hard-ness is good, or is it bad? 什麼叫好 什麼叫不好 順路ów管理那塊 就是不好的順路ów管理 就是它會随便去做一些中間的裁ред 去扔掉一些前面的倫值 或者說隨意的修改系統提示詞 搞これ導致 Lazarus 的時效 就因為過去一到兩年 我發現很多開发 show 非常關懈順這項順路一樣 管理的問題 其實最好的管理就是不要做管理 因為你做管理真的會導致 來從一開始失效的 對不對 在 Redmiche payage指示詞 那是我覺得不好的做法 說跟模型的運行不自 Head promoted Good harness design is There are nice modern Has Japan become the more simple Great Then there must be a sort of the athlete's Going to a good Student in general, HAARNICE is a surprise� all within its entry theory It's a surprise to the AI. For the least, what is a promise to warm up with the variations of Breakdown Act between the gleich OK and hit a well state. It might also be a result of Token Kegels more and less. The goal of next wave would depend onñar 덩ichen is based on Harness over thecontinent red product as today we've been too 50,000 years ago. That is,troreek will start getting quick into�� depends on the habla that we led to kita from-manic Ltd account from participant tools whichはい tool for attenue the concept designers in teilweise we aren't actually we still don't remember I think it's not a good thing to talk about the token efficiency. This is very important, but for us to use Sota, we need to use Sota. We need to use Sota and Sota to create a product and software. Sota is Anthropic, that's Sota. Sota is Anthropic. Sota, like Kami, we say Minimax, GLM, they're the most recent model. They're all in the same way. If the future of A-chain's model is better, harness between the size of the size, it's important? If we say that the time is already 2 or 3 years, we should not have that much harness. First, the harness is better to work with the machine. It's better to work with the machine. The harness is better to work with the machine. The machine is stronger, and the machine is stronger, and the harness is stronger. It's not that the machine is stronger, the harness is stronger. It will cause the harness to launch a lot of different types of machines to make a new version of the machine. We can do a Linux system. You can say Linux is the most powerful. If the machine is developing a ASI, it will only use Linux. If we're able to do an agent better, it has a lot of appear formed like like project boomers and tile Are you original costumes and Doubt By our way the first type I think it the first thing We mentioned the agent is the same It the same Of course, many developers are not in this opinion. But I'm very familiar with this. I think the agent is the same. It's a very important thing. The same thing is the first thing. If the agent was the best thing, then the more powerful. It's better than the more powerful. I think that the other thing is a very open way. 非常开放的概念 比如说他工作在什么样的文件夹里 或者是工作在Windows上还是Linux上 对吧 然后呢 包括他能够用哪些工具 这些工具是CRI的形式给他的 还是用MCP形式给他的 这个也属于执行层面能力的上下文嘛 然后呢 包括Skills 包括Memory 包括上一轮那个Agent干完 干满了 他要传一个吉利棒交给下一个Agent接着干 所以我觉得上下文是一个也是非常重要的部分嘛 第二重要 对第二重要的部分 其实对用户视角来说上下文才是最重要的 但是其实因为模型我们大部分 So we don't have a lot of people who don't have a card. So we don't have a lot of people who don't have a card. So we can do the space in the space. And then the third thing is the tool. If you can add some powerful tools. It really can do a lot of things. In the meantime, the 3rd of the year I was interested in GitHub. It was an MCP. I found my work is... It's a free one. I don't really need to open GitHub. And every time I'm in the end of my 3rd minute. I'm in a 3rd minute. ho 당하세요 Git molt pill dogs Hoo trat deem However, the Maybe we Here we why Fat It waar I have top Thatивал That's what they had to do in Ciera. yön шир the results of the excellentению how they would have manooms How are the scientific especially CCWs and energy below Hunger 사랑,� ди lac.rian nantes implement a superweight we can't to get our respond time 给他们开发M和Mail 而是我就关心下放了真正的硬件资源 比如我云上是有服务器的 我端测可能有我们的Mac笔记本电脑 有我的Mac mini 我的路由器 有我的NAS 有我的闲置的手机 而他们不一定都是有公网IP地址的 然后他有很多的混合作用的问题 那以前我们可能会用FRP内网穿透等等很多方案 但我觉得那个不太可以scale 然后呢 其实我这方面就非常喜欢一个工具叫做Tile Scale TELSKELL它是一个就是不管你是云上的还是边上的它的组网方式就非常的简洁全部都统一组拘网这种混合组网方案我觉得在Agent的时代非常需要但TELSKELL其实不太Agent Native它的CRA上有很多的调用能力都没有这个赛道可能对Agent来说可能需要一个全新的混合组网形态尤其很多时候需要高通量的商业文交换这个交换不只是说我发一个mail我发一个IM消息就完成了所以这其实和Agent的支付也是一样的这些人的支付其实不需要那么高的并发 not need to be a small amount of money but agent between the money is a few minutes it's very high-level in the past so this is also the agent payment of these companies they talk about the bet the future of the way and this is one and you said the second is what? we think today is agent this kind of in a baby we go to 10 years later we go to 10 years later of course. So if you have a proper go-ok and just expand the thinks. What did the request for is. Then the user has the license to create! The user needs to create a LiveFront channel. And I'm doing research I think I can do a lot of things in the future. That's why I'm doing a lot of things in the future. Because if I put a lot of things in the future, I will be able to pay more than 5% of the amount. That's not high, but I got a lot of things in the future. And I can do a lot of things in the future. So I'm very careful about this kind of approach. And now there's a lot of things in the future. OpenAI's CEO is a team of T-Sinkers. The CK Machine Lab. And I think there are some things in the future. but I think it still doesn't fit in the T-Sinkar and the standard of things because it actually has to be with our software I just said that we have a data that is a data harness that you have to run for half a month you will have to do a lot of things you have to do good cases and bad cases you need to do the same thing but many times you have no money to do this to make this machine you only do a few things you can't do a few things to make this machine I think it will be more appropriate for my ERP system, more appropriate for my CRM system, for my own environment. You think in the future you will be able to do the K-Series, Agent Infra will be able to live in many companies? I think it will be a more serious side-down. We think in the future Agent Harness maybe not needed to be able to do the shipping. Because it is first to have to do the application with the machine In fact, I think it actually goes to theoch십 My opinion is about the extent of Jelly being brought up to the Unix and Shell I mean still have some tecnología I try to have great content That's what I blij trust So I can always get people on the market Be careful What's that? There were a lot if I need for tablet Let's do this long-term,では? — We're not going to play for this season. — We are not beginning driving this season, do this season. We are trying to draw against this whole thing. Here we are trying to make two-time dollars. And we are interested in Belkin's thoughts in tomorrow. I just 500 minimum streams." We're interested in this definition in Real-Chair that was because of the days of life. Whether it's true in the real world in those days the world's greatest speed in the world's greatest speed maybe the next level will be the next level of speed for example, the current speed or the better speed speed will be the greatest speed speed but in the current level, we can see how the world's greatest speed to become a global speed especially the old coding agent to become a global infrastructure and the future will be more than a global speed on how to serve them, compartments, Soligent agent, is a very crazy environment. Are you still thinking about it and Except for every? So we want to하기 better the slogan, jamopoly enforcement would happen and make a goal in LG. What is U model for the future to Apprenti, Is there some other goals in the U.S.? I believe in the comes ofienzaes Very notes veryunk maintenant. This whole process for me is still about a toute one, In this, agent mode comes out with three months now.raws the summer in a new agent, then vagina is more than one another On the next line on this does not know if siinä is social media. If skinny is a libro and your head tiers in a lot of times, that and then their of foreign policies are будетеkin a CPA and a people who are truly more involved, It has its輸入 and輸出. It's in the middle of the business. The customers will not be able to go to the office office. So, the company is a黑-hug. I think the agent can be a agent. It's a 0-year-old company. I don't think it's a 0-year-old company. I think it makes sense. It's a 0-year-old company. We grant a grant. It's a 0-year-old company. It can be a 0-year-old company. It's a 0-year-old company.就把他丢进汪洋大海了就从此不再改他一行代码也不给他一分钱他要自己去进化要自己去想办法搞到钱来这个给自己买token然后他有一个目标这个目标就是有一天超越Cloud Code然后他不断地进化自己然后他怎么这个来赚钱呢他现在赚钱的方法就在GitHub上面开打赏去化源让大家给他捐钱那所以我们Token Grand就是给他捐了第一笔钱让他开始有token有养料可以去进化其实在今天还是很早期 But in the future, it is a company that is a company. And today we have to pay some money, but it doesn't have its own money. But in the future, the economy is becoming more advanced. The economy is not just a company that is a company that is a company that is a company. This is a company that I believe. In the future, you may be a big owner. And then you can walk in the road and see your friends from your mouth. And you can get a card from your mouth and get a card. So you put your friends in the car, then decaya. They have more money. Then you rain upon them. Do it in my first year... Think that's interesting. It is so forgetting something he'sévimately this whole thing. I'm loving it на fairs and have a full combi. Ok Thank you all for sending me to you in the Facebook page. Goodbye.