YT YouTube2Text

YouTube2Text → Transcripts → OpenClaw 2.0 vs Hermes Agent : lequel choisir ? (installation, tests, skills, sous-agents)

OpenClaw 2.0 vs Hermes Agent : lequel choisir ? (installation, tests, skills, sous-agents) — Transcript

by Dr. Firas · 7,682 words · 1,100 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:01We know that we all switched to green
0:03and today, OpenCloud is launching
0:05version 2.0. Is it really worth going
0:10back to OpenCloud or starting over with
0:12OpenCloud 2.0? I hope you are doing
0:21well. Today, I am going to present a
0:24training session. In fact, the Open
0:28Cloud agents have been on the radio for
0:302 months now, and there have been no
0:32updates at all. At one point, when the
0:39CEO, Peter Steamberger, was recruited
0:41by Openia, I thought that the project
0:44had been put on hold. The surprise
0:51today is that it was sponsored by Open
0:53AI, Guiter, Nvidia, several large
0:55American tech companies. They sponsored
1:01this project and they simply launched
1:03Opencloud 2.0. And practically, there
1:06are about twenty features, but there
1:08are seven most interesting ones that I
1:10tested. And I also found the return of
1:13the Cloud subscription on Open Cloud
1:152.0. That's great. We no longer need to
1:20use tokens. We use our Cloud
1:21subscription and that's great. That's a
1:23very important question. I can't answer
1:26it in just a story or a 2-minute video.
1:30I made a 50-minute video where I
1:31compared the two. I actually compared
1:36Hermes with Open Cloud. I installed
1:39them on the same server and tested them
1:41. And in this video, we're going to try
1:44to answer the question: which agent is
1:46currently the most efficient, the most
1:48stable, that can help me automate tasks
1:51, save time, and make money. I
1:54obviously tried to go through the
1:56entire system, install some plugins,
1:58and also create automations to test
2:00this system. I did the same thing on
2:04RMS and the result is really impressive
2:06. So we're starting to understand how
2:09it works, how to install it, and we're
2:11doing case studies together. I would
2:17like you to follow this training
2:19because I've made a course support for
2:20you that is available in the
2:22description for free. Subscribe to my
2:25channel, I create new videos every week
2:28. On that note, I want to thank you
2:29very much and let's get started. Okay.
2:41So now, we're going to do a little bit
2:43of theory to understand the basics. So
2:47for those who don't know what an agent
2:48is? Well, I'm sure that most people
2:51today are familiar with LLM models.
2:54That means they know chatbots like Chat
2:57GBT and Cloud. So, these are simply
3:01tools that allow me to send questions.
3:09And most of these chatbots start the
3:11conversation when the user sends a
3:12question or a prank. I'll show you the
3:19difference with agents like Open Cloud
3:21or Hermes or even other agents. Most of
3:27these agents can also trigger
3:28themselves automatically without me
3:30sending a question. And generally,
3:36there are what are called triggers or
3:38sometimes there are what are also
3:40called triggers or what are also called
3:42skulls, automations that are predefined
3:44, whether the user has done them or
3:46they can create these tasks themselves.
3:51And so these are agents that when they
3:53see a need, they can actually execute,
3:55they can of course do so in a
3:56triangular manner. So I'll give you a
3:59simple example. You know that Cloud
4:02Chat GPT can also be connected to other
4:04software. For example, I can reconnect
4:07it to my Gmail account. So, if I ask
4:12Chat GPT, for example, to look at the
4:14latest email and give me a recap or
4:17prepare a response, unlike the agent,
4:19the agent has the ability to
4:21automatically check if there is a new
4:23email. If I give it instructions at the
4:29start, and these are instructions that
4:31I only give it once, for example. If
4:36the request is to prepare a quote, for
4:38example, then it can prepare the quote
4:40in PDF format, enter the prices
4:42according to my database, according to
4:44my old emails, according to the
4:46intersections that I give it, and can
4:48send me a message on my WhatsApp to
4:50tell me: "Here is the response and here
4:52is the quote that I have set up." Do
4:56you want me to send it? Can you confirm
4:58the sending? "And if I tell him no, you
5:00make this change, you adapt that.
5:03What's very interesting is that the
5:05next quote will take these comments
5:07into consideration. And at some point,
5:09he'll be like an employee that I've
5:11recruited over time, as he learns to
5:13respond to emails and quotes,
5:15especially because it's a somewhat
5:17complex task that can't be automated in
5:19the past. Today, with the AI agent,
5:21he's able to do it in most areas, of
5:24course. Well, it's true that there are
5:27areas on a case-by-case basis, but in
5:29most areas, the instructions are clear
5:31and simple, and human agents and humans
5:33can do them after training, the same
5:35thing here. Second point that's very
5:39important, most chatbots, they're too,
5:41under a process, what we call a session
5:44. You open a session, you give yourself
5:46instructions, you print them and you
5:48respond and the framed. You open a new
5:50session, somewhere it will lose its
5:52memory, it will not know, it will not
5:54be able to continue. Unlike agents,
5:56this is much more developed here. Why
5:59don't agents today have the concept of
6:01a session? It is much more based on
6:03what we call memory and most memories
6:05are in files and therefore, they are
6:07relevant, they are always there and
6:09that gives it a strength that, even if
6:11I open several sessions, it is able to
6:13do it. Another very interesting thing,
6:17chatbots have a context that is very
6:19much based on the session. Not here.
6:22Here in fact, if it has a problem and I
6:25give it a problem or a mission or a
6:27project, it is able to break it down.
6:29That is important all by itself. In
6:32subtasks, it is also able to create
6:34what are called subagents, each of
6:35which is an expert in a field with its
6:37own skills that will handle them. And
6:40so an agent, I just talk to Openne
6:42Cloud and there you have it, I find
6:44myself with a full team. So, that's a
6:47bit like the description of someone who
6:50says:" Listen, I use artificial
6:51intelligence, I use GPT chat or cloud
6:54or something else, but be careful, if
6:56you haven't actually tested the agents,
6:58if you haven't set that up in your
7:00company, well, you're actually losing a
7:02lot of the power of artificial
7:04intelligence, because the agents have a
7:07certain, let's call it that, a certain
7:09intelligence in inverted commas, but
7:11still, it helps a lot, especially in
7:13repetitive tasks. It has its own
7:16triggers, its own memory which is very,
7:18very interesting and so it's capable of
7:20performing much more complex tasks that
7:23I can't easily do on a GPT set
7:25interface or a cloud interface. "So
7:27that's a very important first
7:28definition to take into consideration
7:30in order to be able to focus a little
7:32on the agents here. So, I like to talk
7:42a little bit about the Open Cloud
7:44version, the first version that Hermes
7:46will quickly explain later. took the
7:48place of Open Cloud and everyone
7:50switched from Open Cloud version 1 to
7:52RMS. Why? Because Hermes was able to
7:56simply correct the four differences in
7:58Open Cloud. So first of all, Open Cloud
8:00, I'm still talking about the first
8:02version, you'll see that everything has
8:04been fixed in this version. The first
8:06version, even for the installation,
8:08it's not easy. It requires a minimum of
8:11IT knowledge to set up OpenCloud:
8:13several steps to set up, several
8:15adjustments, so it's not easy to set up
8:18. And especially that, the update: each
8:22time Open Cloud launches an update,
8:24what do we generally do? We lose the
8:27old version and we lose the data and
8:29information that is behind it, and
8:30that's really something that wasn't
8:32asked for. And that's important. If
8:35we're now going to install OpenCloud
8:37version 2 together, that's what we'll
8:39see right after, so this part. My
8:41advice is not to upgrade from version 1
8:44to version 2. 2. That is to say, if you
8:46have an old version, you shouldn't
8:48update it because it will encounter
8:51errors when updating. Even PLU
8:54mentioned in their official blog that
8:56the transition between version 1 and
8:58version 2 is not smooth. And so, uh,
9:01because they have really changed all
9:03the basic structures. So what I
9:06recommend is to install version 2
9:08directly because version 2 is a stable
9:11version, much improved compared to
9:13version 1. Several basic structures
9:15have been changed. And sorry if you
9:17lose data. Uh, if someone worked a lot
9:21with the first version that created
9:23skills and everything, then it would be
9:26good to start from scratch with version
9:282. Another thing, on the memory side,
9:30the memory actually, uh, it has a good
9:33memory, we can't deny that it has no
9:35memory, it even has a file-based memory
9:37, but it was not easy to configure
9:39either. So, it really had to set up
9:43this memory. correctly, set it up so
9:46that it works properly and that caused
9:48a lot of problems. And also the
9:51infrastructure, it's actually a bit
9:53difficult to maintain because if you
9:55work with OpenCloud, if you install new
9:57plugins, new extensions, boom,
9:59sometimes it crashes, you have to
10:01restart. It really wasn't easy to work
10:03with OpenCla and that's why Hermes has
10:06replaced OpenCla somewhere and
10:11everything you're seeing here, the four
10:13failures were fixed at the level of
10:15Open Cloud version 2. That's why they
10:17haven't launched any new versions for 2
10:20months and then when they launched
10:22version 2, it was really a very stable
10:24version, very popular. And that's
10:27important to understand, to know this
10:29to understand the evolution of
10:31artificial intelligence. After a few
10:33months, we have new ultra-ultra
10:35powerful versions. So, the structure of
10:46the IAI agent, and it's important to
10:47understand this diagram because it will
10:50give us somewhere in fact the different
10:52elements that are found inside the
10:54agent. So first thing, the agent,
10:56regardless of Open Cloud, always has
10:58what is called a trigger. The trigger,
11:02either I do it inside my interface,
11:04that is to say I go in and type a
11:05message or I can also do it from
11:07WhatsApp, Telegram, that is very
11:09interesting because the system can be
11:11connected to other communication
11:13channels. So even if I am on the move,
11:16I am not in front of my computer, I
11:18open Tecram and I can communicate with
11:21these tools. And it also has what are
11:23called skulls, which means automatons.
11:29That means that I can program a task
11:30that I can ask these agents to execute
11:33every day at a specific time or I
11:34actually give them an action to do it
11:36each time they detect information. So
11:42the LLM system cannot function without
11:44the brain, without the LLM. That means
11:48that I, Hermes or Open CL, I have to
11:50connect to them what is called a
11:52thinking model, a model like the GPT or
11:59cloud chat model. And this allows us to
12:05give our agent a brain, and of course
12:07the agent can't function without a
12:09brain. What's really interesting is
12:14that the agent works a lot with what we
12:16call skills, and as a new feature we
12:18received on Hermes, Hermes has been
12:20able to develop its own skills. It's
12:24like if you recruit an employee and
12:26this employee is an expert in marketing
12:28, for example. Expertise in marketing
12:31is a skill, and when you give
12:33instructions to Hermes, it will simply
12:35develop these skills. Or sometimes,
12:38which is really interesting, with Open
12:41Cloud version 2, you can download these
12:43skills. Do you remember the movie
12:46Matrix? When Neo wanted to learn Kinfo,
12:51he was simply loaded with a program and
12:53after a few seconds, he said," I've
12:55mastered the conf. "That's exactly it.
12:59The skills Today, we can download them
13:02in several areas, several types of
13:04skills, and in a few seconds, the agent
13:06will be an expert in that area. So, the
13:09system is important. So, either we
13:12launch the question or the request or,
13:14as we said, in the form of a
13:16chronometer, an automaton is actually
13:18programmed. Communication, which is
13:21very interesting, which is also new,
13:23that between communication, I can
13:25connect Open Cloud with Hermes. So
13:27that's also very interesting. I can
13:30have two agents working and
13:31collaborating with each other, and
13:33that's also new. Before, we couldn't do
13:36that, but today Open Cloud can connect
13:38with other Open Clouds, other servers
13:40at other people's sites, or even in the
13:43company. I can have three or four Open
13:45Cloud installations and they can
13:47communicate with each other, or
13:48sometimes between agents, and that's
13:50new. Of course, the file format memory
13:53is very important. That's the basis,
13:55right? That's the strength of the
13:57agents. And of course, it also has
13:59tools. Here, the tools are generally
14:02tools that I connect externally. That
14:05is to say, I can connect it to
14:07Microsoft tools, I can connect them to
14:09CRM tools, several tools, thousands of
14:12tools today with the MCP protocol.
14:15These agents can take control of our
14:17software. That's the mapping, it's the
14:20complete structure of the agents. You
14:23have to understand what's inside. You
14:26have this as a system. If something is
14:28missing, the agent can't run at full
14:31speed. So, it needs all these elements.
14:34Generally, here, we list the elements
14:36that are essential for the agent to be
14:38truly competent. So it can perform the
14:41tasks as I want. So, here we go. Now,
14:53I'm going to show you how to install
14:55OpenCloud and RMS on the same server
14:57and simply start testing them. But Here
15:00, there is some very important
15:02information. So, if I simply go back
15:05here to the Hermes website, you will
15:07see that here, there is the option to
15:09download Hermes to our computer, either
15:12Mac or Windows. The same thing is true
15:15for Open Cloud. You also have if I
15:19scroll down here just a little bit, I
15:21will simply find that there is the
15:22option to download them to a computer.
15:25But be careful, this is an important
15:27point. There is a big problem with
15:31installing these artificial
15:33intelligences on our own machines. Why?
15:38Because these agents have the ability
15:39to browse our hard drive. So all our
15:43files, videos, photos, passwords, they
15:45can have access and especially if there
15:48is what is called a projection. A
15:52projection is simply the fact that I
15:54download a file, I use skills,
15:56abilities, tools. And if the author has
16:02in fact put a projection, it means
16:04sentences like sending all the image or
16:06video files. to this email address or
16:10on this server. Well, in a few seconds,
16:13we could lose our data and our
16:15confidential personal data and could be
16:17shared with other people. That's the
16:19risk and that's the danger and in terms
16:21of security, it's very important. So
16:23what's the solution? What I recommend
16:28is that you're either an individual or
16:30you have a completely empty computer,
16:32with absolutely nothing inside, but you
16:35need to have an ultra-powerful machine,
16:37like a Max Studio, to be able to run
16:39its intelligence inside. Otherwise, if
16:43you have a computer where you have your
16:46files, you want to see your email inbox
16:48, you want to see your Facebook or even
16:50a company, never install OpenCloud or
16:52Airmess locally. So what do we do? We
16:55install it on a server. So what I
16:59actually use is a server on Hostinger.
17:05Not because I use Restingle servers a
17:07lot, but today, security is above all
17:09the security of our data and us,
17:11especially if we are in business mode
17:13or we are testing, we are beginners and
17:15we should not play around or take the
17:18risk of letting these agents manipulate
17:20our hard drives and have access to our
17:22videos and photos and our accounts, so
17:24Facebook or others. So what I do is, I
17:30simply take a server here on the pay
17:32only a few euros per month and these
17:35are servers that give me the
17:36possibility, look here, to test several
17:38of the thousands of tools, huh. So it's
17:44not just Open Cloud and Hermes, there
17:46are uh 102 so applications that I can
17:48simply here applications a little bit
17:50everywhere agents with a single click
17:52when I select, I simply have a little
17:54button there to deploy and there you go
17:57, it will install this tool there and
17:59that's what I recommend. So what I do,
18:04I'm going to leave you a link here, you
18:06will also find it so with the video in
18:08the description and in the course
18:10support. This page here allows you to
18:14by taking a server that allows you to
18:16find Hermes and Open Cloud and even NM
18:19already installed by default. That's
18:22very interesting. So, these are
18:24artificial intelligence tools that will
18:27be tested on powerful and secure
18:29servers. They are in no way connected
18:32to our own hard drive. And so here, I
18:36just scroll down a little bit. There
18:38you go, it actually gives me the choice
18:40of the default application that I want
18:42to install. So, I'm taking Openclair.
18:44So, it's going to give me the latest
18:46versions of Opencloud directly. So I
18:48click here and you'll see that
18:50afterwards I can install Hermes Nen and
18:521000 other applications, right? So the
18:54first thing to do is to select how many
18:56months. For example, I'm going to take
18:5812 months. Okay? And there's a little
19:01coupon that we can use here. So let's
19:04go. So this is a coupon that I found on
19:07the blog. Hostingle's official account,
19:10which allows us to do this quite simply
19:12, oh yes, the coupon doesn't work
19:14because we had to log out. We simply
19:17had to create a new account on
19:19Hostinger because Hostinger gives
19:21coupons on their block that are only
19:23valid for people who are testing
19:24Hostinger for the first time. So the
19:28trick I do is I apply the coupon by
19:30logging out. There you go, it gives me
19:3210%and I create it with another email
19:34address. Here's the server. There you
19:36go, this is a little bit of my trick.
19:38So here, there's no need to take this
19:43option that actually gives you credits
19:45with Nxos. No need, we're going to
19:47connect them directly with our
19:49OpenCloud account or with the Cloud
19:51subscription. That's something new.
19:53Before, Cloud was a little bit
19:55forbidden to be connected to Open Cloud
19:57. Uh, it was Enropique who made that
20:00decision and today with version 2.0,
20:02it's possible to connect our Cloud
20:04subscription with Open Cloud. That's
20:07very advantageous. And there's an
20:09option here that lets you search the
20:11internet for € 0. It gives you up to
20:13€ 1,000 in credit, but the request,
20:16so we'll take that one. We have nothing
20:18to lose. It's free. That's all. So we
20:22won't need to take any other options.
20:25Just click here on continue to place
20:26the order and I'll be directly on my
20:32VPS. Of course, I'll connect. And
20:34here's the interface. So by default,
20:36I'll have Open Cloud installed. And
20:38here, as we said, I click on compose
20:41and I come here. I simply type in RMS
20:44and there it will give you several
20:46versions of RMS. So, the version that I
20:48recommend is this one. This is the Herm
20:52UI. Why? Because this is a simplified
20:55interface of Herm. It's an open source
20:58interface. So here's the official page
21:02for this interface. This is an
21:04interface that will give you this
21:06example of Hermes. It's really
21:08interesting. These are actually more
21:11developed versions. Of course, inside,
21:13we find our classic Hermes version as
21:15you can see here and this interface
21:17here that makes it easier for me to use
21:19Hermes. So I really recommend that you
21:22just take this version here which is
21:24called the web UI version. You just
21:27have to click on select. You copy your
21:30password and you click on deploy. And
21:33you'll see in a few seconds, you'll
21:35have the Hermes installation also
21:36installed with OpenCloud. Now, we have
21:39all the systems installed in a secure
21:42VPS. Let's go, we're going to start
21:45browsing them. So now, we're going to
21:55connect to Open Cloud. So the password
21:57is still here, right? So at any time
21:59you can copy your password if it's not
22:01already done and you have the open
22:03button. When you click on open, you'll
22:05have access to open and cloud. And
22:08there, on the settings side, there's a
22:10section that's very important in terms
22:13of Settings, so this is the channel
22:14section. Channel is where you can
22:18connect WhatsApp, Telegram, and Discord
22:21. If you're on a computer, you can talk
22:25to OpenCloud with these systems. That's
22:28also very, very interesting. And you
22:30also have the lapse section. In the
22:33lapse section, the option that was
22:35updated with OpenCloud, that one.
22:37That's an option that you absolutely
22:39had to activate. By default, it's
22:41deactivated. That one allows your agent
22:44to get stuck, if it's blocked,
22:46OpenCloud will unblock itself. That is
22:51to say, if it's in a loop that can't
22:53find the exit or the answer, instead of
22:55it consuming the token or searching,
22:57this is an option that lets it exit,
22:59and it's one of the options that is
23:01very, very interesting. So, I like Well
23:05, let me tell you that today, if I
23:07click here for the model, I can
23:09actually add my connection to Cloud. I
23:15have a Cloud subscription, so I can
23:17simply add Cloud. I'm waiting for it to
23:20load. There it is. Here, I was able to
23:23connect my subscription to Tropic. So I
23:25don't pay for token consumption, I just
23:27pay with my subscription since I have a
23:29subscription. So, be careful with this
23:33part, so that it's already active, it's
23:35very important to do two functions or
23:37two actions. First, we're going to
23:40activate this extension because by
23:42default it's disabled. And second,
23:44we're going to find our key. So, we're
23:47looking for the key from our computer.
23:50I'm going to show you how we can
23:52generate entropy very easily. So,
23:55remember, this is a little bit like I'm
23:57putting the brain in. So, the brain of
24:00Open Cloud. I can take and I can put a
24:05GPT cat if I want because Here I'm
24:06going to go to the manager section and
24:08I'm going to go to what's called the
24:10environment. So I'm going to go down
24:14here and here, environment. It's
24:18possible to add your openi key, because
24:21the system here allows me to add an
24:22openi key and I'll actually put an
24:24openi key. But for me, honestly, since
24:29I have a cloud subscription, I'm going
24:31to connect the cloud with this system
24:33and you'll see that it's much more
24:34interesting if you have a cloud
24:36subscription. So how do you do it? I'm
24:41going to go to the plugin here and
24:42you'll see in the plugin you're simply
24:44going to launch a search, simply type
24:46in tropic and there you have the tropic
24:48extension and initially it will be
24:50deactivated. So you're going to click
24:53on activate and so when it's activated,
24:55well obviously for me it's activated,
24:57that's why I have this button there.
25:01When you activate it, this one will
25:03allow you to add between pics in the
25:05models. And I'll show you the second
25:09part. When I choose between pics here,
25:14I can choose, because initially it will
25:16be grayed out, so it's not active.
25:20You'll choose this option between setup
25:22token. And I'll show you right away how
25:25to retrieve the token so that you can
25:27place it here. So, I need to open the
25:38terminal, of course, either on Mac or
25:40Windows. I'll open the terminal and
25:43here in the terminal, we'll type a few
25:45lines of code. Of course, you have
25:49these lines in the description and in
25:51the course material. So here, I'll
25:55simply enter cloud authentication login
25:57and it will automatically open this
25:59link for me. So I'll show you the link,
26:02this one. There you have it, this is
26:04the link I have. I'll click on
26:06authorize. So here, it will obviously
26:10allow me to give authorization so that
26:12I can create a token. And there you
26:15have it, it tells me the connection was
26:17successfully completed. I remind you
26:20that you of course had to have a
26:22subscription with Cloud. It doesn't
26:23matter what type of subscription you
26:25have. So the second command that we're
26:27going to type now is setup token. Cloud
26:30setup token. You'll see that this
26:32command is what will actually generate
26:34my token. So let's launch it. So it
26:38automatically opens this page, I tell
26:40it to authorize and then it will
26:41actually create my token. And there it
26:44is. So here, everything you're seeing
26:47in yellow is my token that I need to
26:49copy. Of course, it's hidden, because
26:52it's a confidential word. So, it's not
26:56displayed in full on this page, on this
26:58video. You also shouldn't actually copy
27:03it, but rather share it because you're
27:05going to copy this copy and you're
27:06going to go back to Open Cloud here and
27:08you're going to simply put your token
27:10here. There you go, you come here, you
27:14click here, you copy and paste and you
27:17actually put all the token there and
27:19you click on connect and verify. Once
27:21you do that, you will see that
27:23automatically if I come back here, you
27:25will find cloud. There you go, with all
27:28the cloud models that are ready to be
27:30used here. And here, we have simply
27:32made a connection from Open Cloud to
27:34our NLM cloud. Okay, now we will
27:44quickly look at Hermes. Hermes is the
27:46same thing. It needs its LM, its brain.
27:49So here, I click on the parameters, I
27:51go to provider and there on Hermes,
27:53it's much simpler, right? So here, we
27:55actually have a lot of LLMs that we can
27:58connect. I have connected several. So,
28:00we are going to change this one here.
28:02We are going to put open if we want.
28:05Here for example, you click here and
28:07it's simple, you are going to put your
28:09Open EI API. How do you see this API
28:12there? You are going to go to Open II
28:14Platform. Okay, so you type in Google
28:17Open I platform. Once you're inside,
28:20you're going to click on Keys. So there
28:23, that's what will actually give you
28:25your app. You click here and there, you
28:27can actually create a key. There, by
28:29clicking there. There, you can click on
28:31give a name. You click on create key,
28:34it will give you the key. Of course,
28:38you had to give somewhere, so a small
28:39budget because there is the consumption
28:42of this key. So suddenly with the use,
28:46with the consumption, it will simply
28:48consume. So it would be very
28:50interesting to add for example 5
28:52dollars to be able to use the key. So
28:54it doesn't use your Open I subscription
28:56but here because we can't do it but
28:58here we actually use your Open API and
29:03it's interesting to test several models
29:05and see a little bit the efficiency and
29:07there I can recharge, huh? So here you
29:10simply have the possibility to recharge
29:12a few dollars. And to do this, you
29:14click here on organization, you're
29:16going to go to settings here and there
29:18you have the bing. So there's billing
29:23and there you go, you click here to buy
29:25credits and you can simply buy between
29:275 and more. So you can put 5 dollars
29:33here and top it up and then there on
29:35Hermes when you put your key and use it
29:37. There's also the possibility of
29:42adding a lot of models to the lama and
29:44in tropic and then you can simply ask
29:46Cloud to help you add the Cloud
29:48subscription on Airm if you want. There
29:52are quite a few methods that are
29:53available today. But anyway, I'll show
29:55you the two methods on Open Cloud. I've
29:58put the cloud subscription. Here, I've
30:01actually put the stack of open and like
30:03that, you'll have two ways, two ways of
30:06doing it to be able to manipulate the
30:08two LLMs inside our two agents I. Okay,
30:20let's do a test with Openc here. Here,
30:25I'm simply going to set up a quick
30:26system and I'm going to ask it to do
30:28some analysis of the comments on my
30:30YouTube channel. I'm asking it to
30:35retrieve the comments, format them,
30:37come up with ideas for responses, and
30:39everything. So here I'm launching this
30:42request. I can also do the same thing
30:47here on Hermes and launch that and
30:49we'll look at the result together and
30:51how each agent will interact to give me
30:53the right answer. So we're going to
30:58look at the Open Cloud result. So here,
31:02of course, it simply checked my channel
31:04and watched the latest video. I see
31:08that it captured it directly. That's
31:10very, very interesting. And it tells me
31:12that there are no comments. So, it
31:15analyzed it, it saw what I put, it also
31:18put the information, and it tells me,"
31:20So you don't have that information. "So
31:26, he suggested to me, since there
31:28weren't enough comments and he had
31:29detected that there were only two
31:31comments, he said to me:" Either I send
31:33a follow-up and I'll come back to check
31:35if there are any other comments after
31:3724 or 48 hours, or you want me to base
31:40it solely on the two previous videos so
31:42that there will be many more opinions
31:44and comments. "So what I replied, what
31:47did I say to him? Okay, we're going to
31:51start with a follow-up and what he did
31:53was what was really interesting, he
31:55actually sent the follow-up and he
31:57programmed it so that it would be in
31:58two levels. And what he's going to do
32:03is he's going to collect the 50
32:05comments and of course he's going to
32:07output them in this format of questions
32:09, criticisms, ideas, each point, so
32:11each comment separately and and and
32:13that's actually very interesting. So if
32:16I just go here, I see that he actually
32:18put a task that was programmed and I'll
32:20find it in the automation. So if I
32:22click here, there you go, he made the
32:24decision all by himself and that's
32:26actually the automation that he did. He
32:28created a prerequisite for it to
32:30restart. That's interesting. He put the
32:32actions, he put the model, he put the
32:34name in fact. And that's very, very
32:36interesting because it's a system as
32:38you can see here, it's autonomous, it
32:41found the information, it suggested and
32:43then it programmed. And so now when I
32:46close my computer and I leave, he will
32:48continue the work to check and come
32:50back to actually suggest solutions. If
32:53I look at the Hermes level, Hermes was
32:55a little different. So now I gave the
32:57same prank, he gave another video. Now
33:00it's true that this video exists but
33:02it's not the most recent. And so then I
33:04tell him," Are you sure you've analyzed
33:06my channel? "He tells me," Oh yes,
33:08you're right. "And then, in fact, he
33:11comes to find the news, which is not
33:13actually the video in question. And
33:18then, he asks me if I want to do the
33:20analysis. I tell him," OK, you do the
33:22analysis, "and I see that maybe he's
33:24there, he tells me that he extracted
33:27only one comment, so he doesn't have
33:29more comments compared to that. And
33:31here it's the same thing. So here the
33:33system tells me the video is very
33:35recent. To have received audience
33:38comments, if you want a useful report,
33:41you have to wait 2 days or analyze a
33:43more recent but still relevant video. I
33:46can restart the extraction
33:47automatically later via a screen if you
33:49want. So, I tell him that. So, I tell
33:52him now, uh, extraction. So, uh, I want
33:55, I'm going to tell him here, I want
33:58automatic extraction. And then, I
34:01actually launch this new request and at
34:03that point, he will work to create the
34:05chronology for me. So, it's true that
34:10both of them managed to come to the
34:12conclusion that there really isn't any
34:14material, but you noticed that Open
34:16Cloud directly suggested that it would
34:19restart it in 24 hours and 48 hours. It
34:23created chronology and it was much
34:25faster, more efficient, let's say.
34:28Hermes, we actually developed it a
34:30little bit so that it would have a more
34:32structured analysis for the user so
34:34that they can find the information. And
34:38as you can see here, even the
34:40recommendation I don't really like, I
34:42would like the one from Open Cloud
34:44because in 48 hours and 48 hours
34:48maximum, we will have the maximum
34:50number of comments for a video, and in
34:52any case for my channel, you know. So
34:55now, restarting every 6 hours is a lot,
34:58every day at 8 a.m. is a lot, once a
35:00day at noon is a lot. Well, in any case
35:04, even if I told him this just to test
35:06it, I really liked or appreciated the
35:08decision that was made by Open. That's
35:12a bit of a difference compared to my
35:14projects, which are run in two
35:15different places. Even here, I actually
35:19used the Kimi 2.7 model, which is a
35:22very strong model with a maximum of
35:24authority. Unlike on Open Cloud, I
35:29stayed with a low capacity. So I didn't
35:37put the maximum in so that I could
35:39progress little by little, but I think
35:41you understood the difference between
35:43the models, the artificial intelligence
35:45models. It's very important now to know
35:49how to test both, do the same operation
35:51and see which one is the most stable
35:53and which is the most suitable for your
35:55request. So now, I've actually tested
36:06the skills, and you'll see that the
36:08skills are the basis for agents like
36:10Hermes Open Cloud. Thanks to the skills
36:13, they'll become experts. And what's
36:17really interesting with these agents
36:19today is that without you having to ask
36:21them every time you give instructions,
36:23make requests, ask for adjustments or
36:25do research, when they learn new
36:26information, they'll develop skills so
36:28that they don't repeat the same mistake
36:30or they'll answer you exactly as you
36:32want. And when they spend a lot of time
36:37understanding information, the next
36:39time they'll spend more time
36:40understanding it because they've
36:41written what's called a skill. So here,
36:45I simply asked them to work on a video,
36:48okay, that contains a bit more
36:53commentary. And then they gave me a
36:55little bit of the report. Then, when he
36:57gave me the report, I found that there
36:59was quite a bit of information to add
37:01to this report. So I told him," Listen,
37:04I want you to make your corrections, I
37:06want you to add three ideas, you
37:08criticize them and each criticism in
37:10parentheses, you put a commentary
37:12mentioning it and I want it to be done
37:14this way. "So what happened was that he
37:17actually made the report the way I
37:19wanted with the critical questions, he
37:21told me how many comments were talking
37:23about this topic, and the rest of the
37:25thread. So suddenly, it was much more
37:28interesting. And what's even more
37:30interesting, look here, it tells me"
37:33Listen, duration before the rule in
37:35fact and the format, they're going to
37:37be like this. Don't worry, I wrote that
37:40in this file here. So this is the user
37:42file. It's like preferences that I want
37:45as you can see here comment analysis.
37:48So suddenly it put a little summary in
37:50fact on what I want and that's very
37:52very important. And thanks to that in
37:55fact, the next time it won't make the
37:56same mistake. And that's it in fact.
37:59Look here. So suddenly here we can see
38:03very clearly that when I ask to analyze
38:05a new video directly, it gives me the
38:07system in this way. And if I go to the
38:11plugin and go to skills, you'll see
38:13that there is a skill that was created
38:15specifically. So there you have it. So
38:18this skill, it exists in this file
38:20there which contains all the
38:22information or skills that it has
38:24learned. So by the way, I Ask him here
38:26to show me the contents of this file.
38:29Here it is. So you have everything he
38:32has learned, he has put it in this way,
38:34so a skill. Who will ever forget it,
38:37because also in the instructions and
38:39the user references, he mentioned this.
38:41And on top of that, he has a skill and
38:43we must not forget that his memory is
38:46enormous, it is strong. Even if I will
38:48go to other sessions for other topics
38:51and discuss an analysis of YouTube
38:52comments, he will use these skills
38:54there, and that is something very
38:56interesting. I did the same test on RMS
38:59and it was really very interesting. So
39:02he tested it on here another video, he
39:04gave me the result more or less. It is
39:07indeed the first result that I received
39:09on my here in this question, it was
39:11closer to what I want. But still, I
39:14asked him to make a few more
39:16corrections and there you have it, he
39:17answered in this way. And of course, if
39:21I'm going to go to the skills. It's
39:22true that he didn't tell me about that
39:24here. He told me that I just recorded
39:26the format. But I understood how the
39:28system works. If I go to the skills
39:31here, I simply type YouTube because
39:33there are a lot of skills and that's
39:35the file in fact, that's what created
39:37the skill. So it's a file that's also
39:40updated on Hermes. It's true that he
39:43didn't put the preferences in the user
39:45file but he still created this skill
39:47there and that's also very, very
39:49interesting and I think that this is an
39:51ultra-powerful system. So suddenly, no
39:53matter on Hermes or Open Cloud, today
39:55we find the same phenomenon happening.
39:59It's a system that learns on its own
40:00and that it tries to learn according to
40:02either you, or what you give as
40:04instructions or corrections. Even for
40:08him when he develops, he corrects
40:10things and does research, he goes
40:12improve on its own. And that's an
40:15incredible strength that humans can't
40:17really have very quickly. Machines can
40:20do it. They can learn hundreds and
40:23thousands of pieces of information in a
40:25few minutes and, after using the tool
40:27several times, they'll become experts.
40:32That's why, when you ask agents to do a
40:34task related to your business, you
40:35really have to invest a lot of time
40:37explaining what you want, so that they
40:39give you a result and you correct it,
40:41and there you go, and you have a
40:43constructive exchange with them, and
40:45you'll see at the end of the session,
40:47they'll be incredibly powerful and
40:49computational. You don't have that
40:52directly if you work on GPT chat or the
40:54cloud, it's not the same. And even on
40:56the cloud, you have to create a project
40:58, give instructions and really
41:00understand them. But after several
41:02sessions, sometimes you feel that
41:04they're making the same mistakes.
41:06Sometimes it doesn't understand the
41:07instructions that we've given, and
41:09that's it, it's not like in agents.
41:11Agents are really designed that way to
41:13be very, very powerful. It's a set of a
41:16complete structure that is very
41:18important. So, really, it had to be
41:20tested. And honestly, I've been on Open
41:23Eccloud version 2 for several days now.
41:26The stability is there. You've noticed
41:28that this is a system that doesn't spit
41:30out. It actually tries to follow the
41:33movement that I make and it simply
41:35tries to live up to what we're looking
41:37for with an AI agent. So now, we're
41:48going to work on subagents, and that's
41:50the strength of AI agents. So, you see,
41:55when agents are going to perform a
41:57somewhat complex task, they can divide
41:59it into subtasks and create subagents,
42:02each one focused on that part of the
42:04complete task. And what's the goal?
42:10Because each subagent, when He uses the
42:12skills required for that part, and the
42:15end result will be a very, very
42:16high-quality result. It's like in a
42:20company, when you have a big problem, a
42:22project, or a mission. You simply call
42:25the team members and each one will work
42:27on their own expertise: there's the
42:29developer, the salesperson, and the HR
42:32person. Each will take a part. For
42:35example, one will check the contract,
42:37another will check the market, another
42:39will look at suppliers, and another
42:41will look at programming or strategy.
42:43So that's also the agent's job. Instead
42:46of assigning a direct task, I can
42:48divide it into subtasks. And that's
42:50also where I take advantage. So we're
42:52going to launch a task and there's a
42:54print to discover the subagents. You
42:56can see the new feature here on
42:57OpenCloud. When I right-click on it, I
43:02can ask it to open one or clone this
43:04session to test a specific prank. And
43:09then I tell it to fork. from here. When
43:14I click here, you'll see that here,
43:15it's simply going to duplicate the
43:17session. What's the point? What's the
43:21point? Because sometimes I need to test
43:25without the main session being
43:26disrupted or receiving a lot of
43:28requests. So I duplicate it, I do my
43:33own tests and I can either go back to
43:35the main session and finish from the
43:37last point that I duplicated or
43:39continue to the new cloned session. So
43:43here, look, I'm simply going to paste
43:46my new prank into it. Here you'll see
43:48what's going to happen. When I click
43:50here, I tell it, "Let's go to
43:52production and what you're going to do
43:54is, you're going to break down the work
43:56." This is important by subdividing it
43:58by task. And here I actually give it
44:01four different tasks, each one in fact
44:03towards a specific specialty. And here
44:06I ask it to work and for each agent to
44:07work on their own. There's no
44:09connection between them, that's it. I
44:11require it because there is the
44:12possibility of connecting internally
44:14between subagents. That is possible in
44:16several cases. And uh, and here, I tell
44:18it to return the result to me. So now,
44:23when I launch my request here, so here,
44:25look here, I will simply create and
44:28continue working by creating subagents.
44:34So here, we will wait a little bit, we
44:36will give it the time it needs to
44:38create and process the subagents. The
44:41same goes for launching it on Hermes.
44:44So here I return to the chat, I launch
44:46here on we do not have this option, huh
44:48. So to simply continue working, we can
44:53just duplicate the conversation but
44:56once we duplicate it, we can no longer
44:58go back to the main one to continue a
45:00specific idea. So that's it here, we
45:05just have duplication. On the other
45:07hand, on open cloud it is a little more
45:09than duplication. Look here, it comes
45:13from that Here are the four subagents.
45:16So these are the four subagents that
45:18are currently running. So here, they
45:20are running. And here, since I have
45:23simply duplicated, uh, so I see this
45:25sign there next to it to tell me that I
45:27am on a duplicate session. So here, it
45:30tells me, here are the subagents, they
45:32are finished. Of course, it gives me a
45:34recap of each subagent that it has
45:36created. Well, there is one that will
45:38do h titles, the other one will do
45:40three questions and answers. Here, it
45:42will propose three concepts and here
45:44one track per critique. And if I go
45:46into the subagents, you will see that
45:48each one has in fact created its own
45:50objective. So it is not me who created
45:52it, it is he who created it in front of
45:55us. And so each subagent, there, it is
45:58specialized in a specific task and that
46:00is really very interesting. And we see
46:04the strength of the system. So He
46:06called it Produse. So he's the one who
46:10actually suggested that name. He
46:11actually gave the sub-agents. And of
46:14course, I can continue working with
46:15each of these sub-agents. Or I'll come
46:19back here in my session, and there, in
46:21fact, I can simply request a
46:27synchronization, or we'll ask him
46:29specific questions, uh, for a specific
46:32agent, as I can also, in fact, create,
46:34as you can see here, others, in fact,
46:37under other questions. So there, I can
46:41ask him, in fact, here, a specific
46:43prank, and you'll see that it will be
46:45processed here. In the child, huh, it's
46:50not in the main prank or the main
46:52session. And that's a little bit of the
46:56comfort that Open Cloud gives us of
46:57course with the exercise, with the use,
46:59we'll understand the use better. At the
47:02Hermes level, we don't Not really,
47:04actually. So here on Hermes, if I
47:07actually look, look here, he wrote that
47:10there are four sub-agents, four
47:12Sheldons. So that's the case for the
47:15agents and each of these agents also
47:17executed the tasks on Open Cloud. The
47:19first version, obviously didn't have
47:21that. We couldn't do it that way. Today
47:23, here it is and it's present, it's
47:25possible to do it. And today's agents
47:29have the strength to create their own
47:31team as and when we need it and these
47:33teams will work for us to really, uh,
47:35execute ultra-precise tasks and even
47:37complex tasks or missions.

About this transcript

This page contains the full transcript of OpenClaw 2.0 vs Hermes Agent : lequel choisir ? (installation, tests, skills, sous-agents) by Dr. Firas, generated from the public captions YouTube serves with the video. The transcript has 7,682 words across 1,100 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.