---
title: 'Hermes Agent: The Self-Improving AI That Creates Its Own Skills'
source: 'https://youtube.com/watch?v=EHlqRx0r4BI'
video_id: 'EHlqRx0r4BI'
date: 2026-09-03
duration_sec: 1528
channel: 'David Ondrej'
---

# Hermes Agent: The Self-Improving AI That Creates Its Own Skills

> Source: [Hermes Agent: The Self-Improving AI That Creates Its Own Skills](https://youtube.com/watch?v=EHlqRx0r4BI)

## Summary

This video introduces Hermes Agent, an open-source, self-improving AI agent by Nous Research, positioned as an alternative to OpenClaw. The presenter demonstrates how to deploy it on a Hostinger VPS, configure it with various providers and models, and highlights its unique self-improving loop powered by GEPA (Generic Evolution of Prompt Architectures). The video also compares Hermes Agent to OpenClaw and Agent Zero, emphasizing its developer focus and four-layer memory system.

### Key Points

- **Introduction to Self-Improving AI** [00:00] — The era of self-improving AI has begun, with auto-research going viral and now Hermes Agent as a self-improving alternative to OpenClaw. It can create new skills and improve based on mistakes without user intervention.
- **What is Hermes Agent?** [00:27] — Hermes Agent is an open-source agent by Nous Research, combining elements of Claude Code and OpenClaw. It is developer-focused but general-purpose, optimized for open-source models.
- **Deployment Options** [01:10] — Can be deployed on a VPS, GPU cluster, or server-like infrastructure. Supports various models via providers like Moose Portal, OpenRouter, GLM, Kimi, Minimax, and is highly customizable.
- **Installation and VPS Setup** [01:36] — Installation is a single command. Recommended to give the agent its own VPS for full capabilities, analogous to giving a new employee a dedicated computer.
- **Self-Improving Loop** [04:51] — Every ~15 tool calls, Hermes pauses, evaluates itself, and saves learnings as skills. This is powered by GEPA (Generic Evolution of Prompt Architectures), which mutates prompts based on execution traces.
- **Installation Process** [06:30] — Copy the quick install script from GitHub, paste into the Hostinger terminal, and it automatically checks and installs dependencies like Python, Git, Node.js, and ripgrep.
- **Integrations** [08:13] — Supports Telegram, Discord, Slack, WhatsApp, Signal, and CLI. The presenter demonstrates both CLI and Telegram integration.
- **Model Selection** [09:10] — Uses OpenRouter for flexibility. Recommends setting API key limits (e.g., $30) for security. Chooses Claude Opus 4.6 as the best model, but mentions open-source options like Minimax M2.7 and GLM5.
- **Memory Architecture** [15:04] — Hermes Agent uses four distinct memory layers: memory.md and user.md, SQLite session archive, procedural memory (skills), and optional Honshaw user modeling. SQLite is preferred for agent memory due to query efficiency.
- **File System Checkpoints and Intelligent Forgetting** [16:54] — Built-in Git-based checkpoints and worktrees. Intelligent forgetting extracts knowledge before compressing context to avoid losing important information.
- **Comparison with OpenClaw and Agent Zero** [20:19] — Hermes is newest and less polished but improving fast. It has no heartbeat (unlike OpenClaw's 30-minute default). More developer-focused vs OpenClaw's mass-market approach. Can migrate from OpenClaw with 'hermes claw migrate'.
- **Skill Creation Demonstration** [22:31] — The agent creates skills strategically after complex tasks (5+ tool calls) or overcoming errors. It scans available skills before each response and loads matching ones.

### Conclusion

Hermes Agent represents a significant step toward self-improving AI agents, with its unique GEPA-driven loop and four-layer memory system. The presenter predicts that self-improving loops will dominate AI projects through 2026-2027, making it essential for developers to experiment with such tools now.

## Transcript

We have entered the era of self-improving AI. First came auto-research, which went super viral. It's a self-improving agentic loop from Ajay Kapati. And now, there's a new AI agent called Hermes Agent, which is basically a self-improving alternative to OpenClaw.
So imagine if your OpenClaw could create new skills and self-improve based on the mistakes it did, based on the conversation with you, without you having to lift a finger. This is what the future of AI agents will look like, and Hermes Agent is the first taste of it.
So in this video I'm going to show you what Hermes agent actually is, how you can set it up yourself, and how it compares to OpenClaw and Agent Zero. So first off, what is Hermes agent? Well, it's an open source agent by Nose Research.
And imagine if Cloth Code and OpenClaw had a baby, where Hermes agent is still focused on developers, but it's a general purpose agent like OpenClaw. And it's built for developers, but again, anybody can use it, and it's optimized for open source models.
because those research has been releasing over those Hermet models for years. Alright, so here is the official GitHub repository from Hermet's agent. As you can see, it's already at 15,000 stars and this is what the growth chart looks like. So yeah, the project is going viral and if you aren't using Hermet's agent
or at least haven't tried it, you are not on the cutting edge of AI. So the first thing you have to figure out is where do we deploy it, right? And we have multiple options, but as it says here, you can run it on a Fiber VPS, a GPU cluster or a server-like infrastructure
that's called in the North American Idol. And you can use any model as you can see Moose Portal, OpenRouter, GLM, Kimi, Minimax but it really supports any model and I'm gonna show you just how customizable and configurable Hermes Agents really is.
Now the installation is literally a single command and to give it the full powers is best to give it its own computer which means establishing a new VPS where you give the AI Agents its own virtual private server so it can do
anything on that server just like a human would. Like if you hire a new human for a company, you would give them a new computer. You wouldn't have five people at your company share a single computer. And this is the ideology that we need to have for AI agents as well. Now, right now,
the best BPS option for AI agents is definitely WholeFinger. As you can see, they have a step-by-step open call installation, which you don't even have to touch the terminal at all. And as I mentioned later in the video, I'm going to compare Hermes agent to open call, but this KVM2 server can
actually host many different AI agents at once. It can run an agent zero instance and Hermes agent instance open code close code open claw all at once like the ktm2 server is pretty op so this is going to be the first link below the video i'm going to click on deploy and this is going to take
us straight into the whole thing your card so here's for like the time period i highly recommend going for 24 months because you set up your agents once and they run forever they work in your life in your company completing tasks saving you time every single day so there's no point setting up an
agent for a month so go with 34 months and as you can see right now hosting has a very nice discount on the 24-life option. Now, if you want another 10% off, go to the right and click on coupon code and type in the code DAVID for additional 10% off. Oh, and shout out to Holsinger for sponsoring this
video. As you can see, OpenClaw will auto deploy on this VPS, so we're gonna have OpenClaw, so then all that remains is just installing Hermes, which again, super easy, I'm gonna show you that step by step. So you can do that even if you're a complete beginner, location select the closest one
to you. Alright, and let's click on continue, it's gonna take us to the Holsinger cart. Here fill out your billing address details and select the payment you want to use and then click on submit payment. Once the payment goes through you'll go into the Hostinger panel. So on the left click on
VPS and here you'll see all of your different Hostinger servers. It'll probably look like the DocPlay one or the Docker one if you went through the open core install I just showed you, but I'm going to show you how to install Hermes on any VPS, right? So maybe you have an older Hostinger
VPS that we used to use for NA10. Well guess what, I'm going to show you how to do that. So just click on manage whatever VPS you have and then we need to click on the terminal in the top right. This is the integrated Hostinger terminal which is the easiest way to manage a VPS and actually before
we install the Hermes info script we need to do one thing add user Hermes just type in password something super secure like admin now I'm just joking but you know whatever you can skip that you can enter through all of these and then if the information is correct why boom and then click
Click clear and then type in usermod-A-G sudo hermes. Boom, and we're set up. We created a new user for Hermes agent. And that's how easy it was to get ready because the next step is this one line quick install script right here.
And this is how easy it is to set up a new VPS server for AI agents. So again, use the hostinger install. It's like so easy to set up an open floor instance, a Hermes instance with the hostinger VPS. And again, it's going to be the first link below the video.
And make sure to use code David. Alright, so before we run the quick install, let me explain to you the main thing that makes Hermes Asian popular and the reason it went super viral. I mean, growth like this reminds me of the early days of OpenClaw.
It is the self-improving loop, right? Every 15 calls, roughly, 15 tool calls, Hermes pauses, evaluates itself, evaluates the messages that it will happen, and saves learnings as skills. And this is thanks to the JEPA, or G-E-P-A, I don't know the pronunciation of this, self-evolution.
This is a new concept and basically it mutates its own prompts from the execution traces. But I think this is worth looking deeper into because this is a new invention from news research. So, GEPA or GEPA stands for Generic Evolution of Prompt Architectures.
It's a research pipeline that reads the agent's execution traces. So, aka results of tool calls, what it did, you know, why did that. To understand why a task failed, then it uses an evolutionary or genetic approach
to mutate and improve Hermes' own prongs, skills and code. So whatever could have been the reason for that failure, it goes back, tries to figure that out, kind of like a back-propagate through the chat history,
and tries to adjust some of its prongs. It's literally how the transformer works, but with prongs and code. Yeah, I mean, that's actually a really good analogy, now that I think about it. And this allows the agent to optimize itself over time for the following error,
rather than the user having to tune it. So if you want to get an agent up to speed as fast as possible, Hermes is probably going to be that, right? It's still probably going to be the fully configured Agent Zero. Because, like, I've seen what people can do with Agent Zero, and some of these deployments are insane.
So I not going to say it more powerful than Agent Zero but like from scratch if you don have the technical capability if you don have the use cases and the setup Hermes Ancient will do a lot of that for you And that because it gets better the more you use it There no fine tuning needed no crazy prompt engineering needed The agent takes care of that itself
All right, I want to show you the installation process because again, all it takes is copying this install script, which again, it's on the GitHub repo of Hermes Agents, which I'm going to link below the video. So just scroll down below and into the readme file and it's right here.
quick install copy this boom one click copy then let's go back into hosting your panel and again in your hosting your panel go to the top right click on the terminal this is the integrated terminal that makes it super easy to interact with your vps and just paste it in boom hit enter and
this is going to start the hermets agent installer it's going to check your system so it detected linux with ubuntu and then it checks some other dependencies like uv which is a faster package manager then pick, checks whether you have Python, it wasn't found so it installed it,
and you can see it's taken care of everything. Git, it found that there's Git, Node.js, Node.js wasn't found so it installed it, else checking, ripgrab, fast file search, sfmpack, it's for text-to-speech but it's also for editing files, so it's installing all the useful
dependencies, and this is why I would highly recommend you set up on a DPS, right, this is not a, just a small suggestion. To get the most out of these AI agents, you want to give them a full operating system, a full computer, and the VPS is not just that.
Another thing is that you don't have to worry about these agents messing up anything on your own machine, right? If you give it a new dedicated VPS, like the one I'm using for Boltinger, it can do whatever it wants to be as effective and as powerful as possible.
And again, the analogy to hiring a new employee, but not giving him a new computer, that's a great analogy, you know? Today I'm on fire with these analogies. By the way, Hermes agent also has all these integrations, right? If we do Ctrl F, we can say like Telegram, Discord, Slack, WhatsApp, Signal and CLI.
All of these are possible inside of Hermes agent. I'm going to show you both the CLI and Telegram in this video. So if you're worried about the integrations, if you're like, oh, I'm an open claw, I don't know if I want a fridge. Well, that's not an excuse because Hermes agent has all the things that open claw has basically.
All right, so the install script is still running. Boom. And there it is. So now we have to press enter. And the first step of the onboarding is to select the internet provider. So we have many options. Open router, news portal, codecs.
Such as opening a codecs you can use your chatgbt subscription, which a lot of people don't realize that. It's also possible in open globe but the setup is harder. But here you can just authenticate with your chatgbt subscription and you know, let's be honest, who isn't paying for chatgbt guys?
Hopefully all of my viewers have at least some paid AI plan. But as you can see, it's heavily focused on the free open source providers. Most of these are Chinese ones, of course. But look how anthropic is all the way below these open source ones, right?
So, that's not an accident. We're going to use OpenRouter. Because with OpenRouter, you can use any model, right? So, if you go to OpenRouter.ai, go to the top right, boom, click on settings. So, here I'm going to click on create in the top right to create a new API key.
I'm going to name it Hermes Agent Hostinger. Give it some limits. Let's do $30. It's a good practice to put some limits on API keys so that if it gets leaked or if AI agent goes crazy, he cannot drain your bank account. So again, keep your API keys private. I'm going to revoke mine before uploading this video. Do not share your API keys with anybody. So click on copy. Let's head back to the terminal and do command V and enter. Now the key will not display. So don't be scared when you paste it in. It's not showing. That's how the terminal handles security.
So just make sure to do Command V if you're on Windows, Control V to copy. Now, in terms of the model, again, let's choose the best model possible, which right now I think it is OpenSource 6.
But there's rumors that Netropic has new models cooking, you know, so maybe 4.5 and OpenSource 5 are just around the corner. But, you know, if you just select the best model for Netropic or OpenAI, you can go wrong with that. But if you want to go with some of the best open source models, then currently Minimax
M2.7 is super strong, GLM5 and GLM5.1 are super strong, obviously KimiK 2.5 is classic, but we're gonna go with Opus. In terms of text and speech provider, we can keep the current edge, or if you want to use
OpenAI or 11 Labs, obviously there is an option for that. Next up, terminal backend, I'm gonna keep the current enter. You can just enter through a lot of these. don't enter through this one, this is the agent iteration right, so maximum two
calling iterations per conversation. I would recommend going to 150 with this right, not 60, it's like a minimum or whatever, default is 90 but hey let's push Hermes agent to its limit, let's go 150. Two progress modes, so off, new,
all, verbose, off goes all, context compression, so this is like at what point you wanted to start compressing right, so if you use cloud code or codecs they compress when they hit their limit, which is like 0.95 or 0.99. 0.5 is the default,
that's like half of the context window. I think that's a bit passive, I would recommend something like 0.7 personally. Session decimal, you can leave this default, inactivity and daily reset, that's good, enter that. Inactivity timeout, hit enter. Daily reset hours, so 4 am local time,
that's good, enter. Set up telegram bot, I'm going to go with yes, so type in y and enter. and we need to paste in the Telegram bot token. That way we can interact with Hermes agent through Telegram.
So for this we need to open the Telegram app, the desktop app. If you don't have it, just download it. Or if you want to use a different provider, there is going to be a next step for that. Here we need to go to the bot father. To make sure it's the official one with 8 million monthly users, this is the bot father, right?
So what we need to do is we need to do slash new bot status. We need to name it so we'll say Hermes Hostinger VPS. Boom! Now we need to choose a name so I'm gonna say Subscribe Right Now Hermes Bot.
If you're watching this and if you want more videos on the cutting edge agents such as implementation make sure to subscribe. It completely free guys and takes like two and a half seconds and it helps out a lot and yeah a lot of people are avoiding these cutting edge AI agents because it takes a lot of time to figure out you know first you need to learn where it is how to set up all that stuff how it
converts to open cloud so if you want me to make more videos like this please subscribe it's greatly appreciated so we have this ID the token the telegram bot token just copy this part only go back into hosting terminal and paste it in
and hit enter. Telegram token safe, beautiful, allow user IDs, cross-border, open access. So actually you can get your own user ID as well, I'm gonna show you that. Let's go back in telegram, you need to look for the user info bot, this one, and you type it
within and it gives your user info. So in this user info bot any message you send is gonna basically give you your user ID. I'm gonna blur that, so let's switch to back to the bot folder, boom, and then paste in your user ID, that way only this
is allowed. Okay, next up, use your user ID as the default home channel, so this is where like cronjobs will be messaged to you, and you know, where the agent will message you by default, so yes, you do want to use that, so type in Y, enter, and next you can figure out if you want
to set up Discord, so I'm not going to do that, Slack, no, Matrix, no, Mattermost, no, WhatsApp, if you want WhatsApp, just type in Y and follow the instructions, but again, you can use any
webhooks also and of the main messaging platforms now install the gator is a system the service yes we do want that and we want it to be system service so that is as the most power and most permissions basically which user now this is important that we created the user at the
beginning do not skip that step so we're just going to do her mess and start the service now yes why boom so tools we have all these tools so we have web search and scraping browser automation Terminal Processes, File Operations, Core Execution, Vision, Image Analysis, Image Generation,
Mixture of Agents, Text to Speech, some of these require API keys, there are options that don't require API keys, there are Skills, Task Planning, Memory, Session Search, Clarifying Questions, Task Delegation, Cronjobs, Arnold Training, Home Assistant, yeah there is a
lot, I'm gonna hit enter. In terms of the provider, I would use BrowserEase here, I think it's gonna be the most powerful, hit enter and we can just use BrowseList API key. Now before we do that, let me explain a bit more
about how Hermes Agent works. So the memory and storage is quite different than OpenClaw, which OpenClaw is full of markdown files, but inside of Hermes Agent, it's focused on SQLite persistent memory. But it's a bit more sophisticated than that.
In fact, Hermes Agent has four distinct memory layers. So first it has the memory.md and user.md, similar to OpenClaw, but then it has SQLite session archive, Which, this is like a paradigm where you have to understand the data should be stored not in the way that's most comfortable to you, the human,
but in a way that's most comfortable for the agent. And SQL is literally designed to query large databases. If you know the history of data formats and databases, that's where SQL was invented. As a way to easily and quickly query large amounts of data, or in a custom way, data from large databases in any way you want.
So, SQLite is amazing for agents. You can ask any agent, say like, would you prefer CSV files, Markdown files, SQL? The smartest agents will all tell you SQL. So even if you don't understand it, please give your agents access and ability to work with SQL databases.
Trust me, this will greatly improve the memory of your agents. So if you're struggling with your agents having bad memory, which a lot of people in open cloud complain about the memory, switching to SQL databases on top of Markdown files is definitely the right way. Now their skills, their procedural memory, right?
the agent creates them for the task to be complete and they self-improve during use. So that's the fourth way, it has a four distinct memory layer. And there's the Honshaw user modeling, which is optional. And this is a dialectic chat where the agent asks, what does the user care about?
And gets a reasoned answer. And yeah, so four distinct memory layers inside of RealizeAgent. It also has a built-in file system checkpoints before the structure of operations. So it utilizes basically Git in this way, without you having to tell it, you know, do
a git commit, stage this to git, or create a new branch, nothing like that. And in fact, it also has git worktrees built in. So I just made a video on that, so if you're not familiar with git worktrees, and you want to learn more about them, watch this video right after this one.
Also, Hermann's agent has intelligent forgetting, so it extracts the knowledge before compressing the context. So basically it saves the most important knowledge from the chat history before running the compression, so it's not lost.
All right, now back to the setup. So, connection closed. okay we're gonna have to restart the setup if this happens to you i'm gonna show you what to do go back to your whole finger um panel go to the terminal again boom top right you know after a
long time it times out again typing clear the terminal and type in her max set up boom we can do the quick setup for the missing items only uh uh uh where were we seven enter you can also use
File AI for API keys. Uh uh uh. We're gonna skip a lot of these. Here's Tavili for AI web search, Firecrawl for web search and scraping. But
the main thing is the browser use. I'm gonna hit enter now. And let's go to browseruse.com to get the API key. Boom. Click on get started. Just speed around the sign up. I'm gonna do GitHub. Here's the API key. So they give you super quickly.
So I'm gonna copy that. Boom. Back to hosting the terminal. Paste it in. Enter. messaging platforms, we're good, we already have Telegram and we're good to go. So we can start the gateway. So to chat with it in the terminal user interface
just type in Hermes. So you can begin chatting it with the TUI or the CLI. There we go. Boom! Here we have Hermes agent with the logo, with Cloud Open 4.6, with all the
available skills and all the available tools and let's test the message if it And you can see the model and there it is. So it's working. We're using Hermes agent directly in the terminal.
And you can see how many tokens is the system prompt and you know plus my short prompt here and the percentage of token window and the exact time that passed in class message Very nice observability But I also going to show you how to start it inside of Telegram which means we need to start the gateway So click on Ctrl C a few times clear boom and type in Hermes Gateway Start
Alright, so if you type in Hermes Gateway, it's going to start the gateway, and now we should be able to chat with this bot. So let's go back to Telegram and click on this bot, click on Start.
see if we can chat with him okay formulating typing typing let's see does Hermes agent work in telegram here there we go hi David welcome on Hermes area assistance I'm here to help you
with all sorts of things amazing so you can just chat with it normally what skills do you have built in so it's a long response here's one of my skills autonomous AI agents so it's built in
Close code corrects open code Hermes agents, it can spawn sub-agents by itself for coding tasks, lots of creative skills, data science, yeah there's a lot of different skills. Did you see Minecraft? Minecraft modpack servers, yeah, a lot of different skills for 80 plus different building skills and again, the main thing is, as you use it, it creates new skills from your conversation.
Now the big questions a lot of you are asking for is, David, how does it compare to H0? How does it compare to OpenCore? Well, first of all, Hermes is the newest project, right? Agent Zero is the most mature, nearly two years old, so Hermes is going to have the most bugs,
it's going to have the most unpolished things. If you saw earlier, the terminal user interface, it's not as polished as Cloud Code, as Codex, as OpenCode. It's still rough around the edges, but the project is new.
It's improving fast, so you have to give it some leeway. Also, if you use OpenCode, you can just run Hermes, Cloud, Migrate in the terminal. and by the way, as I mentioned, the hosting of the VPS can easily handle multiple agents.
So if you have a VPS already and you have OpenClaw there, if you type in these three words, Hermes Claw Migrate, it's going to migrate your setup from OpenClaw into Hermes agents.
You're not starting from scratch. Very clever strategy, if I must say. But it has no heartbeat, right? In OpenClaw, it has a heartbeat, I think, every 30 minutes by default. In Hermes, there is no heartbeat.
And overall, I would just say Hermes agent is more developer focused, versus OpenClaw is more mass market. But I mean, if you look at these tools, it's not all developers. As I mentioned, it's really general. A lot of it is like software development, Red Demink, you know, ML, AI, QA, data science.
So it is more developer focused, but it's general. Like anybody can use it, anybody can find value from it. So even if you're not a developer, I wouldn't discard it. I would definitely try it. And that's the approach you should have anytime a new, innovative AI tool like this comes out.
Dedicate, sit down for two hours and try it. Yeah, some of these tools you're going to have to use case for, but the worst thing you can do is dismiss them. In 2026, you really cannot afford to not be AI first. You have to be AI first and you have to push whatever it takes to be on the cutting edge.
I don't care if you have 10 kids and 2 jobs and running 3 businesses. Elon Musk literally is that person. He runs 5 businesses with 14 kids and he's still on the cutting edge because he's trying
to see people about AI. But listen, anybody can block out 2 hours per week to try a new AI tool for themselves. And again, you can watch videos endlessly but like at some point you have to try it. So let's discuss this ability to create new skills.
Do you have the ability to create new skills and if so, how exactly does that work? And explain the self-improving loop you have built, built in not built. There it is, you can see the gateways musing aka thinking. I want to see what's in the system prompt of this.
Alright, so it's very verbose by default by the way. So I'm gonna say, please update your memory.nd to be a lot more concise. I don't like verbose responses.
Yeah, let's see how good it is at self-updating Marginal files. Boom, memory plus user prefers concise, non-verbose. Show me exactly how you update a typefile and why.
Alright, so let's look at this. I can create, update, delete skills, see how it works, creating skills. When I successfully complete a complex task, usually five plus two calls, overcome tricky errors, or discover. Okay, so it doesn't create skills randomly.
Only when it makes sense. You know, if you're on a single topic for a long time, it will create a skill out of that. If there is an error that you're battling with and it figures it out, it creates a skill out of that. So it's very strategic when it creates skills.
And any time you tell it, remember how to do X, and it will package it as a skill. And here's the self-improving loop. before every response I scan all available skills to see if one matches the current task, if it matches it loads it. Okay, so it's basically describing how skills work, so that's nothing crazy.
Now let's talk about the integrations briefly, because this is what made the open code go viral. So obviously Hermetician has all these as well, Discord, Telegram, WhatsApp, Slack, all the main integrations you want. And by the way, we're adding these to Agent Zero as well, so...
There is some that are already released, like Telegram and email. WhatsApp is coming in the next few days. In the last 3 days, I think, on agent0 we've released more updates than in the last 2 months combined. I mean, not as good in the works, so credit to agent0 developers as well.
But, you know, this video on Herbets agent. So, also, the backend is OpenAI compatible, so you can use it inside of OpenCore as a tool, if you want. You can just have Herbets agent running in a CLI in your development pipeline for deployment, right?
Because it's so versatile, you can use it in many different ways. In fact, there are six terminal backups, local, docker, sshd, tonax, invert and model. So, it runs literally anywhere, anyhow, whether it's for your own personal use case or for managing servers in a data center.
So, yeah, this is Hermite agent and we'll see how this goes. Obviously, there is a lot of alternatives to OpenClaw as of late, but this one has a very unique take by making it a self-improving agent.
And my prediction is that for the rest of 2026 or 2027, all major AI projects will be about this self-improving loop. Right? We've just seen the tip of the iceberg. Auto research and Hermes agent, this is just like a little taste of what's coming, guys.
So, if you want to prepare yourself for the rest of 2026 or 2027, go watch this video right here, which is about auto research, which I uploaded a few days ago. So, it's right here. Go watch that now.
It will explain auto research by Andrzej Klopathy.