贝联珠贯创始人兼 CEO、前阿里巴巴集团研究员毕玄，以《以 SREAgent 为例：看垂类 Agent 的竞争力》为题作主题分享。
SREAgent，是一款面向企业 IT 系统的 AI 运维助手，通过分析监控数据和日志，帮助运维人员排查故障、定位原因并处理问题。 毕玄以 SREAgent 的实践为例，讨论了一个垂类 Agent 公司绕不开的问题：大模型越来越强，客户为什么还需要垂类 Agent？
毕玄坦言，他们团队曾认为，让 Agent 长时间执行任务、熟练使用通用工具，是拉开竞争差距的关键；但随着大模型进步，这些能力正逐步成为模型本身的能力。在他看来，给大模型“补短板”，可以带来短期优势，却撑不起垂类 Agent 公司的长期竞争力。
机会藏在具体业务的“脏活累活”里。以运维为例，每家企业的监控系统、数据来源和采集方式都不一样，如何接入并用好这些数据，需要逐一解决。
他介绍，SREAgent 3.0 已尝试让客户用自然语言描述需求，由系统生成相应的 Skill，降低数据接入门槛，但这种功能也容易被同行追上。 更难复制的是背后的专业经验：面对大量日志，哪些信息有用、该如何处理，又该怎样结合客户业务，建立反映系统运行情况的模型？这套模型还要能随着业务系统的变化持续更新，否则仍离不开人工维护。
真正支撑长期竞争力的，是对具体业务的理解，以及围绕专有数据形成的采集、处理、业务建模和模型持续迭代能力。
毕玄强调，尤其是模型能否随业务变化自我演进，将成为垂类 Agent 公司拉开差距的关键。而最终的检验标准，是 Agent 能否把问题解决好，运行是否可靠，行为是否受控。
以下为毕玄讲演的完整速记，模力工场在不改变原意的前提下进行了订正和编辑。
Agent 壁垒变了，从长程任务、通用工具，到拼业务模型自我进化
谢谢 Kevin，确实认识很多年了。当年我还在淘宝的时候，我的老板费青，应该是他劝说 Kevin 把 InfoQ 引入中国，在中国有了第一次 QCon。那个时候，QCon 对整个中国互联网公司的帮助，我觉得是非常大的。因为那个时候，我们很多技术都是看国外互联网公司做了些什么。所以我觉得，InfoQ 这么多年来举办的各种会议，对整个技术圈的帮助是非常大的。
今天是 1024，其实我现在已经很少在外面做技术类分享，因为我自己已经不大管技术了。创业已经 5 年了，我们现在更多在看生意或者经营层面的事情。今天是 1024，我还是很希望回到一个技术的角色，分享一下我们对技术发展的一些看法。
从去年开始，我们就认为，所谓 AI Agent 公司不大可能是 5 年前诞生的。真正能够成为 AI Agent 的公司，多数都是这两年诞生的。因为整个 AI 也就是这几年的事情，前年开始，AI Agent 在业内就已经比较火了。今年，大众的整体感受应该越来越强烈，因为个人 Agent、办公 Agent 开始大面积流行。
大家从广告层面感受到的变化最明显。去年的广告更多是大模型，今年的广告很多已经变成了 WorkBuddy，或者大洋彼岸的“muse”之类的。Agent 已经开始成为大家很强的一个共识。
很多人会越来越觉得，Agent 这个方向肯定是重点。但是，对于所有的个人、公司来讲，大家的空间还有多少？能做什么？像我们做垂类 Agent 的公司，不断被各种人问到：做垂类 Agent 公司的竞争力到底是什么？所以，我今天只是想以我们公司的产品为例，讲一下我们看到的垂类 Agent 竞争力的变化。
但这其实是一个动态过程，因为 AI 的发展变化特别快。包括我们去年在一些会上做的分享，后来再看，一两个月后就被打脸，而且打得很重。这就是因为 AI 发展太快了。所以，今天我讲的话，很有可能明天就被打脸，这很正常，没有关系。
无论对做技术的人，还是做业务的人来讲，最重要的是有自己的判断力，而不是你的判断正确与否。正确与否没有那么重要，但如果没有自己的判断力，就会有很大的问题，就会什么火就做什么，但其实不一定应该这样。
我以我们公司的产品为例。不过，我也不希望这里变成一个广告大会，所以会非常少地讲到我们公司的产品。
先说一下我们的感受。我们是去年上半年开始做 Agent 的，是一家做运维类 Agent 的公司。运维可以认为是一个很垂直的领域，跟研发一样，研发是一个大的垂直领域，运维是另外一个大的垂直领域。
当时，做 Agent 的公司是怎么做的？大家看，也就一年多一点，这个变化是非常大的。
一年多以前，我们开始做 Agent 的时候，要做的是自己盖整栋大楼。如果把做 Agent 看成盖一栋大楼，最开始，Agent 公司除了不用自己做大模型以外，大模型之上的所有东西，都要自己来盖。
如果回头看，大家会看到，当时很多 Agent 公司都在做长程任务执行。包括现在，大模型公司也会讲，基于这个大模型可以跑一个 24 小时的任务，甚至一周以上的任务。怎么样做到这一点，是大模型公司一直在研究的。
在大模型公司之前没有空研究这个的时候，谁在研究？不就是 Agent 公司在研究吗？
所以，那个时候，大家会在网上看到很多做得非常好的 Agent 公司的技术分享。比如，我们一直说，必须感谢像 Manus 这样的公司，因为他们分享了很多在长程任务执行过程中，Agent 公司应该做好什么，比如解决好大模型的幻觉、上下文等各种问题。
Anthropic 也分享了很多。他们的博客还是很值得看的，尽管有别的问题，但博客很值得看，他们会分享长程任务等应该怎么做。所以，当时在长程任务执行上，大家都会看这些公司怎么做，去学习他们。我们也要在这上面做很多东西。
除了保证一个任务能长时间执行以外，对于 Agent 来讲，还需要在大模型的基础上，更好地使用各种工具，因为只有工具才能让大模型变得强大。
比如，Manus 以前用 VM Use 的方法，做了一个很好的抽象：人在数字世界里，比大模型更强的一点，是你拥有一台电脑。所以他们做了一个抽象，叫 VM Use。当然，后来还有更多，比如浏览器，以及 Computer Use 之类的，现在已经进步很大了。
在这个基础之上，才是客户自己的专有数据，以及客户自己的专有知识。
所以大家可以看到，去年上半年，一家 Agent 公司要做的事情非常多，投入也会非常大。这就决定了，可能这一块做好了，别的就不一定能做好。但在这个过程中，这栋楼一直在变化。
今年大家已经非常明显地感受到，大模型公司开始在长程任务执行上取得很大的进步。比如我现在用 Codex，基本上扔一个任务给它，最重要的是前期跟它探讨得非常清楚。聊好以后，就可以告诉它：“你一次性给我做完，整个过程中不要问我任何问题，你自己决定。”
基本上就是扔给它，让它跑 12 个小时，甚至 20 个小时。有可能唯一的问题就是 Token 不够用，经常要重置。我有一个晚上用了三张重置卡，才跑完那个任务。
所以，现在我觉得，大模型在长程任务执行上已经做得非常好了，很少需要 Agent 本身再做太多事情。
第二点是工具层面。大家可以看到，大模型使用工具的能力在大幅演进。看海外公司每次发布大模型，都会不断讲这些。比如更早一段时间，OpenAI 发布时重点强调，用 Codex 操作 Blender 之类的 3D 软件，怎么更好地建模。这些都是在向大家展示，大模型本身使用通用工具的能力越来越好了。
所以，对我们来讲，在长程任务执行和通用工具层面，Agent 公司未来只是做一些弥补：哪里还有一点做得不好，就弥补一下。但我们相信，总有一天，大模型公司会做得非常好。
我们公司以前有个误判。去年下半年，大概 10 月份的时候，我们认为，一家 Agent 公司非常强的竞争力就在这两层：通用工具和长程任务执行。这是 A 公司能比 B 公司做得更好的重要核心竞争力。
如果看当时的 Coding Agent，这一点非常明显。好的 Coding Agent，一定是在这两点上做得非常好；差一些的，就是这两点做得差一点。但今年到现在，我们已经完全不这么认为了。
我们觉得，对于垂类 Agent 公司来讲，这两个绝对不会是你的核心竞争力。这两个以后一定是大模型公司自己的核心竞争力，其他人我们认为没有什么机会。这些事情，大模型自己做好就可以了，其他公司就等着它们做好，或者现在帮它们弥补一下。
短期是有作用的，因为能在短期拉开差距，这也是一个重要的点。但长远来看，不要把这个作为重点。
作为垂类公司，最麻烦的问题是什么？工具可以通用，但大家如果仔细研究，就会发现，很多垂类 Agent 要解决的问题，所需要的工具没有办法通用。
以运维为例，这一点比较明显。每家公司的监控数据、各类数据的采集方法，不可能全部通用。当然，有些通用，但大多数都不通用。这意味着，大模型本身没有办法做好一个通用工具，垂类公司必须在上面自己做一层。比如，有些会通过自己写 Skill 等方式来做这一层。
所以，我们认为，垂类 Agent 公司要找到那些大模型不能天然做好的工具。如果它天然就能做好，那这个垂类 Agent 公司可能要想一下，你到底还有什么价值？
像 PPT 这种东西，我们认为会非常通用。以前，大家可能觉得 OpenAI 做 PPT 做得很糟糕，现在不会这么认为了。像我的整个 PPT，其实全部都是 Codex 做的，根本不是我自己写的。我只是跟它讨论了一下想法，里面所有的图、文字，没有一行是我写的，我连改都没改过。以前我还改，现在觉得已经不需要改了，做得真的挺好。
所以，在通用层面，我们认为大模型自己一定会做好。垂类公司一定要去找那些不能通用的工具。
当然，大家会觉得，这好像很定制化。但我们认为，未来就是很定制化的事情才有足够的价值。这是对创业公司来讲的，大厂是另外一种思维。
对小公司来讲，我们觉得，你需要找到这些脏活累活，并且愿意把它们干好。所以，第一个是找到一些垂直场景，以及这些场景里的垂直工具。
像 SREAgent，我们做 3.0 版本的时候，跟过往最大的不同是，以前很多工具仍然会让客户自己写。现在很多产品也是这样，比如用一个 Agent 平台，平台会告诉你，在上面写 Skill，就可以完成专有数据、知识的采集等工作。
但大家想一想，对于很多客户来讲，他们仍然会认为这件事情很有门槛。因为写 Skill 跟写代码，说实话，没有太大差别，甚至还要另外学习。就算是程序员，也得另外学 Skill，这是另一种知识。
而且，Skill 如果写得不好，影响可能非常大，因为你可能会误导模型。我们以前经常讲，Skill 对整个模型的泛化能力有很强的引导性，如果做不好，反而会导致效果更差。所以，怎么写 Skill 也是一个很大的问题。
以前我们也觉得，客户不就可以自己看、自己写吗，很简单。但今年做 3.0 的时候，我们认为，这也是需要突破的。所以，现在在我们的产品里面，如果客户要对接自己的数据来源，其实不需要自己写 Skill，在上面用自然语言描述就可以。
当然，我们无非是把这段自然语言用 AI Coding 翻译成 Skill。但不同的是，我们在里面植入了自己的思想，让写出来的 Skill 能够保持模型的泛化能力，同时做好整个数据采集等工作。
所以，做了 3.0 以后，我们认为，未来的软件将会具备自进化能力。
我们相信，以后大部分软件都是这样。现在大家用的很多软件是不能改的，以后你只需要跟它说，软件展现出来的内容就可以被改掉。不需要你自己写，也不需要你自己用 AI Coding，因为 AI Coding 并不是完全没有门槛。
这是一方面，但我们也不认为它会成为壁垒。因为我们看到，在云栖大会前发布 3.0 以后，某些友商也同样拥有了这个能力。只要我们有这个想法，并把它展现出来，自然有人会做出一样的东西。这里的壁垒仅在于 Skill 怎么写好。
大家可以看到，这是一个演进的过程。对垂类公司来讲，另外一个很强的壁垒是，比如运维会面对很多日志，采集之后，到底怎么处理这些日志？这是一个非常复杂的话题，需要专业知识背景。
一段很长的日志，到底哪些地方有用？如果直接丢给大模型，通常很容易出现各种问题：可能这次是对的，下次是错的；或者效率不高，Token 消耗很多，等等。
所以，采集到数据以后，很重要的是对数据本身做处理。这是我们认为垂类公司非常重要的壁垒之一。但最好是，处理这些数据的知识不是公开的，因为如果是公开的，大模型一样可以学好。
因此，大家需要看，你做的垂类 Agent，一是数据采集，二是采集后的数据处理，这些部分是不是比较私有化的东西。
第三个，是在这个基础上的建模。采集完数据、处理完数据，怎么在上面建立一套垂类模型？像 Palantir 也会把它叫作“本体论”，但我们觉得，“模型”这个说法可能还是比较适合。
对于我们来讲，在运维上非常重要的是，采集数据、建立数据模型，再结合客户的业务，把它做成一个客户的运行模型。这才是我们认为垂类 Agent 最重要的壁垒。
大家可以想一想，这套思想可以套用到其他方面。如果是通用 Agent，比如做 PPT，数据采集、基于数据的建模，包括工具，都是非常通用的。这意味着垂类公司很难做好，因为会面对非常大的竞争。
建完模以后，我们后来发现，还有更重要的一步。我们认为，这是各家垂类公司会拉开巨大差距的地方：今天建的模型，明天还能不能用？这个模型必须具备不断自我演化和迭代的能力。
比如对我们来讲，今天模型建好了，明天系统变了，这个模型还能不能适用？是不是还要人去维护？如果还要，这个东西就不可能有用。所以，模型本身必须能够自我学习、不断迭代下去。
当然，在大模型层面会提供一些基础能力，比如“CANNES”，可能会提供一些记忆、迭代演进的能力，但到底怎么做，还得自己做好。
我们认为，一个 Agent，最重要的第一点是高质量地解决问题，第二点是可靠，第三点是受控。受控是现在越来越强调的一点。
未来的部分我就没有写了，因为对我们来讲，一切仍然有很强的不确定性。
但我们现在认为，做 Agent 公司的核心，都在于思考数据采集、数据处理、数据建模，以及模型迭代。这是最关键的四个层面。我们相信，这四个层面会成为垂类 Agent 公司非常强的竞争力和区别点。
接下来，大家肯定会看到，模型厂商不断希望往上移，因为上面才是最大的价值。越往上，越贴近业务，价值越大。所以，未来是一个互相进步的过程。
对于 Agent 公司，肯定要思考一个问题：模型不断进步的时候，客户为什么最后还要用你，而不是直接用大模型？
就像我做 PPT，以前还用别的东西，现在觉得没有必要了。所以，我觉得所有做公司的人都需要思考这个问题。
以前我跟 ChatGPT 讨论时，它说，你应该想一下：“如果 ChatGPT 进步了 10 倍，为什么客户还要用你？”因为它真的有可能进步 10 倍，这就是现在的状况。
我们最近在搞一个活动，这是唯一的广告，哈哈，请允许这个广告停留一分钟。用 SIA 解决你的问题，可以赚 1 万块钱。大家可以扫描一下二维码。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Pixon, Founder and CEO and former researcher of the Alibaba Group, United Kingdom of Great Britain and Northern Ireland, shared under the theme “In the case of SREAgent: Looking at the Competitiveness of Agent”.
SREAgent, an AI transport assistant for the enterprise IT system, helps transport personnel to sort out failures, locate reasons and address problems by analysing surveillance data and logs. By taking as an example the practice of SREAgent, it was discussed that there was an implausible Agent company: LLM was getting stronger, and why did the client still need an Adgent?
As the LLM advances, these capabilities are becoming the capacity of the model itself. In his view, the LLM “compensation board” could offer short-term advantages but could not sustain the long-term competitiveness of the Agent company.
Opportunities are hidden in the “dirty work” of specific operations. For example, each enterprise has different monitoring systems, data sources and collection methods, and how to access and use these data needs to be addressed on a case-by-case basis.
He described SREAgent 3.0 as an attempt to allow clients to describe demand in their own language, to generate the corresponding Skill by the system and to lower the data access threshold, but this feature is also easy to follow by peers. What is more difficult to replicate is the professional experience behind it: in the face of a large number of journals, what information is useful, how should it be handled, and how can it be combined with client operations to create a model that reflects how the system works? The model will also need to be continuously updated as the business systems evolve, otherwise manual maintenance will remain indispensable.
What really underpins long-term competitiveness is an understanding of the specific business and the ability to gather, process, model and model continuously iterative around proprietary data.
It was emphasized that, in particular, the ability of the model to evolve itself in line with business changes would be the key to opening the gap for aging Agent. And the ultimate test is whether Agent is able to solve the problem, whether it works properly and whether it is controlled.
The following is a complete summary of the lecture given by Bixon, which was revised and edited without changing its original intentions.
Agent barriers have changed, ranging from long-range missions, common tools, to self-evolving business models.
Thank you, Kevin. It's been years. When I was hunting for treasures, my boss, Fizheng, should have persuaded Kevin to introduce InfoQ into China for the first time in China. At that time, QCon helped the entire Chinese Internet company, and I felt very big. Because at that time, many of our technologies were looking at what foreign Internet companies did. So I think that InfoQ's meetings over the years have been very helpful to the entire technology community.
Today is 1024, and I don't really have much of a technology sharing outside, because I don't really have the technology anymore. It's been five years since business started, and we're looking at business or business. Today is 1024, and I would like to return to the role of technology and share some of our views on technological development.
Since last year, we've thought that the so-called AI Agent is unlikely to have been born five years ago. Most of the companies that really became AI Agent were born in those two years. Because the whole AI, that's what happened in the last few years, has been pretty hot in the industry since before. This year, the general public should feel more and more, because the individual Agent, the office Agent, is starting to get popular.
The changes felt most clearly at the level of advertising. Last year's ads were more LLM, and this year's ads have become WorkBuddy, or “muse” across the ocean. Agent has begun to become a strong consensus.
A lot of people are going to be thinking, "Agent, this is definitely the point." But for all individuals and companies, how much space do we have? What can we do? Agent companies like us are constantly asked by a variety of people: what is the competitiveness of Agent companies? So, I just wanted to take the product of our company as an example of what we see as the evolution of Agent's competitiveness.
But it's actually a dynamic process, because AI's evolution is particularly rapid. Including what we did at some meetings last year, and then look at it, it's hit in the face in a month or two and it's hard. That's because AI's going too fast. So what I said today, it's very likely that I'll be hit in the face tomorrow, which is normal and doesn't matter.
The most important thing for those who do technology and those who do business is to have their own judgment, not your judgment. Right or not is less important, but without its own judgement there will be great problems, and there will be no fire, but this is not necessarily the case.
I take as an example the products of our company. However, I do not want this to become an advertising convention, so very little will be said about the products of our company.
Let's start with how we feel. We started doing Agent in the first half of the year. We're a company that runs Agent. Transport can be considered to be a very vertical area, and as with R & D, R & D is a large vertical area, and transport is another large vertical area.
At the time, how did the company that did Agent do it? As you can see, just a little over a year, this is a very big change.
More than a year ago, when we started doing Agent, we were going to build the whole building. If you look at it as building a building, at the beginning, Agent will build everything on LLM, except for LLM.
If you look back, you'll see that a lot of Agent companies were doing long-range assignments. Even now, LLM would say that based on this LLM can run a 24-hour mission, even for more than a week. How to do that is what LLM has been working on.
Who was working on this before LLM? Isn't it Agent?
So at that time, you'll see a lot of very good Agent technology sharing online. For example, we've been saying that we have to thank companies like Manus, because they share a lot of what Agent should do in the long-range mission, like solving problems like LLM hallucinations, context.
Anthropic shares a lot. Their blogs are worth looking at, but despite other problems, they are worth looking at, and they share long-range missions and so on. So at the time, in long-range missions, everyone would see what these companies did and learn from them. We have a lot to do up here.
In addition to ensuring that a mission can be carried out over a long period of time, there is a need for Agent to make better use of tools based on LLM, which alone can make LLM strong.
Manus, for example, used the VM Use method to make a good abstraction: you own a computer in the digital world, a little more than LLM. So they made an abstract, called VM Use. And then, of course, there was more, like a browser and a Computer Use, and now there's a lot of progress.
On this basis, it is the client's own data and the client's own know-how.
So as you can see, in the first half of last year, there was a lot of work to be done by an Agent company, and there was a lot of input. That's the decision. Maybe this one's done, and the rest isn't. In the process, however, the building has been changing.
It has become very clear this year that LLM is beginning to make considerable progress in the implementation of its long-range mission. For example, I'm now using Cordex, basically throwing a mission at it, and most importantly, talking to it very clearly. When you're done, you can tell it: "You finish it all at once, don't ask me any questions during the whole process, you decide."
Basically just throw it and let it run 12 hours, even 20 hours. Maybe the only problem is that Token's not enough, and it's always reset. I used three reset cards one night to finish the job.
So now, I think LLM has done very well in long-range mission execution, and rarely needs Agent to do much more of himself.
The second point is at the tool level. As you can see, the ability of LLM to use tools is evolving significantly. The overseas companies keep saying this every time they look at publish LLM. For example, earlier, Open AI publish focused on how best to model using the Codex to run a 3D software like Blender. This is a demonstration that LLM itself is becoming better equipped to use generic tools.
So, for us, at the level of long-range mission execution and common tools, Agent is just going to make up for it: where there's one thing that's not right, make up for it. But we believe that one day LLM will do very well.
Our company had a miscarriage of justice. And in the second half of last year, about October, we thought that one Agent was very competitive on both levels: common tools and long-range missions. This is the A Company that can do better than B Company's important core competitiveness.
If you look at Coding Agen, it's pretty obvious. All right, Coding Agen must have done very well on those two points; that's the difference between them. But so far this year we do not think so at all.
And we think that these two are definitely not your core competitiveness for the underclass Agent. These two must be LLM's own core competitiveness, and others think there's no chance. These things, LLM does it on its own, and the other companies are waiting for them to do it, or now help them make up.
The short-term is useful, as it opens the gap in the short term, which is also an important point. In the long term, however, this should not be the focus.
What's the most troublesome question as a proprietor? Tools can be universal, but if you look at them carefully, you will find that many of the impurities Agent's problems are not universal.
This is evident, for example, in terms of luck. Monitoring data for each company, methods of collecting data of various kinds, cannot be used in their entirety. Of course, some are common, but most are not. This means that LLM itself has no way of developing a common tool, and that proxies must do their own floors on it. For example, some do this by writing their own Skill, etc.
So, we think that the Drifting Agent is going to find the tools that LLM can't make. If it's natural to do it, then this rare Agent company might think, what are you worth?
Something like that, we thought would be very common. Before, you might think Open AI did a bad thing with PPT, and now you don't. Like my whole PPT, it's all Codex, not me. I was just talking to him about all the drawings, all the words, none of the lines I wrote, and I didn't even change them. I used to change, and now I feel like I don't need to. It's really good.
So, at the general level, we believe that LLM will do it by itself. Prosperous companies must look for tools that are not universal.
Of course, you'll think it's quite customized. But we believe that the future is a very customized thing to be of sufficient value. This is for start-up companies, where large plants are another kind of thinking.
For small companies, we think you need to find these dirty jobs and be willing to do them well. So the first is to find some vertical scenes and the vertical tools in these scenes.
Like SREAgent, when we do version 3.0, the biggest difference with the past is that many of the tools used to be left to the client. Many of the products are now the same, for example, using an Agent platform, which tells you that writing Skill on it will complete the collection of proprietary data, knowledge, etc.
But think about it. For many clients, they still think this is a very high threshold. Because writing Skill and writing code, to be honest, doesn't make a big difference, even learn something else. Even a programmer has to learn Skill, which is another knowledge.
And, Skill, if it's not written well, it could have a really big impact, because you could be misleading. We used to say that Skyll's ability to extend the entire model was very guiding and, if not, could lead to even worse results. So how to write Skill is also a big problem.
We used to think that it was easy for the client to read and write. But this year does 3.0, and we think it's going to need a breakthrough. So, now in our product, if the client is to cross-reference his data sources, he doesn't really have to write Skill himself, just in natural language.
Of course, all we have to do is translate this natural language into Skill. But the difference is that we've implanted our own ideas in it, so that the Skill that's written can keep the model broad-based, and do the whole data collection thing.
So after doing 3.0, we think that the future software will be self-evolutionary.
We believe that most of the software will be like this. Now you can't change a lot of the software you're using, and then you just have to tell it that what it shows can be changed. You don't have to write for yourself, and you don't need to use AI Coding because AI Coding isn't completely without a threshold.
That is one side, but we do not think that it will become a barrier. Because we see that after publish 3.0 in front of the Yunus Congress, some of our friends have the same ability. As long as we have this idea and show it, it is natural that someone will do the same thing. The barriers here are only on how Skill writes.
As you can see, this is an evolutionary process. Another strong barrier for proxies is, for example, how do we deal with logs when they are collected? This is a very complex topic that requires a professional background.
A long log. Where does it work? If you throw it directly to LLM, it is usually easy to have problems: maybe this time right, next time wrong; or if it's inefficient, Token consumes a lot, etc.
Therefore, when data are collected, it is important to process the data themselves. This is one of the barriers that we consider to be very important in the case of proxies. However, it is best that the knowledge to process these data is not public, as LLM can learn as well if it is public.
So, you need to see, Agent, you do data collection and data processing after collection, which are not more privatized.
The third is modelling on that basis. How do you build a pristine model of data collected and processed? Like Palantir would call it "theory," but we think the term "model" might be more appropriate.
For us, it is very important to gather data, build data models, then integrate them into a client's business and make them a business model. That's what we think is the most important barrier to aging Agent.
You can think of something else to use. In the case of universal Agent, such as PPT, data collection, data-based modelling, including tools, are very common. This means that proxies are hard to do because they face very high competition.
After the modelling was completed, we later found that there was a more important step. In our view, this is a place where the big gaps can be opened up by the leading companies: The model we built today, will it work tomorrow? The model must be self-evolving and iterative.
For us, for example, today's model has been built, tomorrow's system has changed. Can this model work? Do you want someone to maintain it? If you want to, it won't work. The model itself must, therefore, be self-learning and evolving.
Of course, at the LLM level, there are some basic capabilities, such as CANNES, that may provide some memory, and an iterative evolution, but what to do is to do.
In our view, the first, most important, point is a high-quality solution, the second is reliable and the third is controlled. Controlling is an increasingly important point.
I have not written the part of the future, because for us there is still a great deal of uncertainty.
But we now believe that the core of Agent is to think about data collection, data processing, data modelling and modelling. This is the four most critical dimensions. We believe that these four dimensions will become a very strong competitive and distinguishing point for Agent.
Next, you will certainly see that model manufacturers are constantly hoping to move up, because that is the best value. The higher the level, the closer the business is, the greater the value. So the future is a process of mutual progress.
For Agent, there must be a question to think about: Why does the client end up using you instead of using LLM when the model is advancing?
It's like I made PPT, used to use something else, and now I don't think it's necessary. So, I think all the people who do this need to think about it.
When I was talking to ChatGPT, it said, "If ChatGPT is 10 times better, why use you?" Because it could really be 10 times better, that's where it is.
We have recently launched a campaign, which is the only commercial. With the SIA to solve your problems, you can make $10,000. You can scan the two-dimensional code.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.