So Hermes agent just released mixture of agents and this is a new feature that allows us to get frontier level intelligence even without having access to fable or GPD 5.6. Now you might be thinking but David what even is mixture of agents? Well let me explain. Mixture of agents consults several different AI models right Grog GPD Gemini opus whatever you want. Then it sends it to a single aggregator which usually is the most powerful one and only then it takes an action and sends you an answer. So basically you mix multiple different agents could be from different providers from different companies to get the best out of each model. Now the magic of this is that it actually defeats the best available models publicly even GBD 5.5 and Opus 4.8 which currently are the top models available to us. You know the the slaves mixture of agents is better than both of these. So in this video we'll go exactly how to set up Hermes agent with mixture of agents enabled. I'll explain what this is how it works and when you can actually use it. Now the reason why this matters is because the biggest problem in the entire AI industry right now companies have stopped releasing their best models. I know it sounds crazy for the past 5 years this would be absolutely insane because every company wanted to release their best model to be the one in the lead but right now both OpenAI andropic have very powerful AI models that are not publicly available. So we the open source community we need to take matters into our own hands and figure out how we can get smarter and smarter AI and get better results out of AI without better models from these closed AI companies and ideas like mixture of agents is exactly what we need. So thanks to MOA we can surpass the capabilities of Obus 4.8 and GBD 5.5 by cleverly orchestrating multiple different models in a way that is productive and useful. Plus, in Hermes, the preset shows up as a model, so it doesn't mess up anything. The tool calling works, memory works, session context works, everything works the same. It's as if you were swapping to a smarter, more powerful model. Now, what's nice about how Hermes did this is that you select which models and which providers you want to use. So, for example, you want to use GPD 5.5 for your CHIGD subscription, you can absolutely do that. You want to use DeepS through Open Router, you can absolutely do that. So you can use your existing subscriptions to run this at no extra cost. Okay. But when should you actually use mixture of agents? Because for simple questions like what's the day-to-day or how many calories did I already consume? This is absolutely overkill, right? So you want to use it for the difficult stuff. Hard debugging, code review, planning the codebase architecture, security hardening. You really want to use it where you don't mind spending more money for the very best answers that AI can give you. Now what you need to understand about mixture of agents is that it comes with presets. So you can have multiple different mixture of agent presets, right? One of them for adding new features, one of them for code review. It's entirely up to you. For example, Grock models are great at Twitter research, right? But you probably don't want to use Grog for serious coding, at least not right now. Elon rumored their Grock 4.5 is going to be released soon. Maybe it's going to be great, but you know, currently Grock is not as good as GPD or Opus. But luckily, this is not a problem because you can have multiple different presets. And don't worry, I'm going to show you all of that later in this video. Now important clarification that I have to make is that mixture of Asians should not be confused with mixture of experts. Okay, these are two similar sounding terms that are completely different thing. Mixture of experts is a architecture. Instead of having like a one large model, it's different experts. So maybe you have eight different experts and you know one of them is more about math, one of them is more about physics, one of them is more about coding. that allows you to have a large model that normally would require a more powerful computer running on a smaller computer because only certain amount of parameters are active, right? Mixture of agents feature that we're talking about. This is several separate AI models that are complete. Each answer your prompt and then it goes to a single aggregator, single powerful model that combines it to a final response. So these are two similar sounding terms, but they're completely different. Now, there are a few more things that you have to know about mixture of agents that most headlines or, you know, most influencers will skip over. First, more models means more tool calls, more tokens. You will pay more and it's taking longer. Okay? So, this is not for minimizing costs or for doing quick work. This is for squeezing the best and the most out of publicly available AI models. Next, your chrome cache stays intact. So, all of the prone caching savings are still relevant here. Also, all of the models that are reference models, they don't debate between each other. They run in parallel, completely independent, and they just submit their responses to the aggregator. Okay. Okay, so the first step is to set up Hermes agent. And the best way to do that is on a VPS. That way you give the agent an entire computer. It can act upon. It can run 24/7 running all of your automations, chrome jobs, workflows, anything you have autonomously even when your own computer is turned off. So what I use for all of my virtual private servers is Hostinger. This is what everyone on my team uses as well. So it's going to be the first link below the video. And also they have dedicated Hermes agent preset that makes the setup super easy. Let me show you. So when you get to the landing page again, first link below the video, click on choose plan. This will show you the different plans they have. The KVM2 one is more than enough. Allows you to run multiple different agents on a single server. Hermes agent, Pi, Open Claw, Codeex, whatever you want to run, the KVM2 plan is enough. Next, select the period. I recommend going for 24 months because not only do you get the biggest discount, but also in the future, all of us will be managing dozens of different virtual private servers. So, might as well just get it now. But to get the best deal possible, go to the right and click on have a coupon code and type in code David. Hostinger was kind enough to sponsor this video. So if you use the coupon code David, you'll get additional 10% off. You can leave this unchecked as well. You can leave this unchecked as well. This as well. Server location, just choose whatever is closest to you. Operating system, keep it as Hermes agent. Or if you want a plain operating system, go with Ubanu. This will take you to the checkout where you have to fill out your credit card details and your billing info. And that's it. After you pay, you'll need to set a root password to your VPS. So, just click on generate and make sure to save this somewhere because we will need this later to connect to our VPS. You can keep this disabled and click finish setup. All right. So, the VPS has finished setting up. So, I'm going to click manage VPS. And here we can see all the overview, all the stats, everything. But what matters is the SSH right now. Don't get overwhelmed. Doesn't matter if you've never used SSH before. We're going to use AI agents to do this for us. So the first step is to open any terminal on your computer and that could be CMAX or that could be the default back OS terminal. Doesn't matter. I prefer CMAX. I would recommend it. It's free and it's really really good. You can easily create multiple different panes and launch different agents in them. All of that, but it doesn't really matter. Just follow along with any terminal. And again, if you don't have a VPS, click the first link below the video to set up one on Hostinger. Again, this is what I use for every single virtual private server me and my team use. Everything is running on Hostinger. And make sure to use the code David for additional 10% off. All right, so back to CMAX here. I'm going to launch cloth code. I have an alias to just launch it quickly. This is the really key using an agent that you already have to set up a new agent, a new VPS. This is complete paradigm. Instead of trying to do everything from scratch yourself, trying to understand long terminal outputs and DevOps and Linux and everything, of course, you should learn the basic terminal commands. You should learn the fundamentals of software engineering. Absolutely agree. But to be as most efficient as possible, you need to be an efficient orchestrator of AI agents. And to do that, you should use agents for as much as possible. So, I'm going to show you how I would do this to quickly and efficiently set up a new VPS with Hermes agent in a matter of a few minutes. Okay? So, just follow along. You can absolutely do this even if you're not a programmer. So, back to Hostinger Panel. I'm going to copy the SSH access and I'm actually going to uh say, tell me where we are on my MacBook. You can talk to these agents in plain English. And just because it's cloth code or codeex doesn't mean these agents are only usable for code. They can do anything on their computer. They can interact with the terminal which means they can launch applications, organize folders. They can download files, rerender them in different file formats, stuff like that. Right? So it analyzed, okay, we're in the home folder on Mac OS, which is not ideal. So what I'm going to do is I'm going to kill it with Ctrl + C. I'm going to CD into documents/David. I'm going to do CD video folder. Boom. In here, I'm going to make a new directory named mixture of agents for this video, right? Boom. I'm going to cd into this and launch the agent in here. That way, it's in a clear project, right? So, it's creating files in here and it's working in here. I'm going to separate it. Boom. We can see that the new CMAX pane launched here as well. I'm going to say createenv in this empty repo and first save the IP address of our VPS in there. Okay, back to Hostinger panel. I'm going to copy this and I'm going to save it like this. Ideally, you would not send this to agents, but uh it's not the worst thing in the world. Okay, so it save that. So, you don't want to do this for your password. It's fine if you do it for the IP, I think. So, now say in the other CMAX pane in this CMAX workspace, try SSH into the VPS. I will fill in the password myself. Okay, so again, it SSH into VPS. Now it's asking for a password. So I'm just going to put in the root password for the VPS that we saved earlier. Boom. We are in. Okay. So I'm say ls just to list it out. Say pwd root. Say check the cmax pane right now. We should be in. And by the way, if you forgot the root password, just goes back to hostinger panel and click on uh change right here. It's super easy to change. It takes 5 seconds. Very convenient. Let's see. Okay. Cloud code said yes, we're in. Okay. uh just run a few terminal commands to learn more about the VPS. All right. So instead of me managing everything on the VPS, Cloud Code is going to do that. In CMAX, it can easily control other PES, right? This is why I would recommend using Cmax over the default terminal on Mac OS because it's not a multipplexer. It doesn't allow you to easily, you know, run multiple agents. If I want to do this and I want to launch a droid here, I can do that very easily. If I want to split this and I want to launch a PI agent here, I can do that super quickly, right? I made a full video on it, but if you want to install it, just go to cmax.com. It's open source and it's free. There's no reason you should be using the default terminal. So, let's see. It learned a bit more about this. So, I was going to say create a readme.md file in this repo and put the most essential info about this VPS in there. All right. So, now we need to get active and we need to set up uh Hermes agent. Usually, it's set up inside of a Docker container. That's how hosting likes to do it. Personally, I like to set it up on root level. So I'm just going to say to cloud code browse the web to learn how to set up Hermes agent on a Ubuntu VPS and then install it root level on my VPS the one we are sshed into. Do all the steps yourself. Enter. Boom. In the meantime, we're going to go to open router to get API key so that we can use any model we want inside of Hermes agent. Again, you can power it for your JPD subscription. You can use uh Versell AI gateway. That's very underrated especially if you want to use GLM 5.2. Versel has a great AI gateway for that. Open router is just a classic. They have a lot of different providers for each model. So if you go like GLM you can see that they have like 15 almost 20 different providers, right? If you go to up time you can see that at all times at least one is active. So go to open router.ai go to top right create a account. If you don't have it just takes 20 seconds to create it. Go to credits charge up couple dollars. You don't need $300. Just do $5 to $10. That should be enough. Then go to API keys and top right click on new and name it mixture of Asians. Make sure to put a limit though because this feature can get expensive. As I mentioned, it uses multiple different models in parallel. So always put a limit here. I'm going to click create and let's see if the setup is ready. Okay, it's still setting up peration. The nice thing is I can zoom out this CMAX pane, right? So this one can be zoomed out and very small while this one cloth code can be larger. So you know the agent that you're talking to that is explaining things to you, you can zoom it in a bit more so it's larger and the other paints can be smaller so you can run multiple of them in parallel and you know have like a bigger observability stuff like that. But anyways this is uh installing all of the dependencies and cloud code is monitoring it right. So anytime there's extra step that it can do itself it will do itself. All of this info like obviously right now it's just packages but like all this info in the terminal when you run terminal command you would have to understand it yourself and that takes a lot of time for an large language model for an AI agent it doesn't take a lot of time you can put in 10,000 tokens 50,000 tokens and it will understand it in a few seconds right that is the difference so when it's running these terminal commands and learning about the VPS and debugging it can understand that like it can input information and digest it way faster a lot of you would still try to SSH into the VPS itself self and try to install her mess and try to understand what's happening. Why is not working? Why is it different than what I'm showing in the video there is so many potential variables that could go wrong. All right, so this is finished running. Seems like it took 5 minutes 29 seconds for cloud code to set this up, which is not bad at all. So next, we need to set up agent so that it's powered by open router. So I'm going to send this prompt into cloud code. By the way, all of this, everything I'm showing you in this video, all the prompts, configs, skills, presets, all of that is available in the second link below the video completely for free. So, make sure to get it and you can just copy paste it to your own agent and it will set up everything for you just like cloth code is doing for me right here. So, I already know it's not going to find the opener API key on this VPS because it's a new VPS. So, let me go back and let me copy this API key created here. Boom. CMAX. So I'm going to interrupt it both of them and do Hermes clear Hermes setup quick setup. Actually no, we don't want to use news portal here. So her setup we're going to do full setup and we're going to select the provider. That's the first step. Let me zoom in a bit. Boom. Open router. It's second right here. Going to paste in the API key for the model. Let's use GM 5.2. Keep current local. You can just enter through this to be honest. And that's that. So we can launch Hermes. And let's send a test prompt. Who are you? And GLM 5.2 is probably the best open source model right now. So Hermes agent is probably the strongest open source agent right now. And there we go. We have Hermes agent running on the VPS cloth code set up 95% of it for us. We just needed to give open router API key so that the Hermes can run. And now let's set up the mixture of agents feature. Next, I'm going to tell Cloud Code to send Hermes a test prompt just to make sure that it can interact with it. I just want to connect the two agents together so that close code can be an effective manager of Hermes agent when we're setting up the mixture of agents preset. Oh yeah, and by the way, if you want all of my skills, all of that is in my skills repository which is in my Twitter bio. So just go to Twitter, it's linked below, and click this davidra.com/skills. You'll get a list of all of my private skills. It's updated literally every single day. I update this all the time. And I have GitHub actions that automatically sync it. So you always get the latest of my personal agent skills. Anyways, uh let's see it. Okay, it responded. So yeah, now Cloud Code can manage Hermes. So let's set up the mixture of agents. I'm going to send another prompt to cloud code to configure a mixture of agents preset in Hermes. And these are the four reference models, right? So this is going to be the references GLM 5.2, GPT 5.5, Kimik 2.7 code, and Opus 4.8 all through OPI router. And the aggregator, the main model that's going to be making the decisions is Opus 4.8. So all of these are from open hour. And again, these prompts, you don't have to retype them. All of this is in the second link below the video in these presets. So you can just copy paste it. Get everything from this video. Completely free. It's going to be emailed to you. Second link below the video. All right, let's go back to Cmax and let's see what's happening. Close code is taking time. It's going to send a Okay, it closed the session. So it can even exit out of it, right? It can operate the terminal better than me and you. So why would you try to operate it yourself? Just describe your intent. Describe what you're doing. Describe your goal to these agents in plain English in the most high level way you can and let the agent do the manual work. The defining skill of the AI era is the ability to ask better questions. It's to have better logic. It's to think more abstract rather than saying do this task, do this task. Why are we doing this task? What is the broader goal we're trying to achieve? Is it to build a great startup? Is it to solve a disease? Is it to get more customers? Is it to have a better onboarding experience? Is it to hire a new person to grow the business to, you know, automate parts of the like you need to go one or two levels higher up than a specific skill? Not like browse this website. You know, give it the objective you're trying to do. I'm right now hiring for multiple positions. By the way, link below as well. Instead of saying like, okay, next I need you to find people in this area. I'm saying like, okay, this is the ideal candidate. This is the criteria we're doing. I'm starting a slash goal either in Codex or in Hermes and it's doing everything for me. This is why it pays off to be locked in on the cutting edge of AI. The leverage you get on your time is truly incredible and it cannot be matched. And it's not just being better developer or being better speaker writer. It's all of these skills combined. You need to be more technical. You need to be better at verbalizing your thoughts and you need to ask better questions. So cloth has ran for three and a half minutes and it should be done. Let's see what's happening here. Okay, this is too long. So, I'm just going to do SL short. It's a custom CL command to just make the answer simpler and shorter. When you're finding yourself repeating the same prompts over and over, just create a preset. Yes, write it now. Okay. I just wanted to confirm this layout for me. Looks good. And you can even set the temperature. Look at this. Reference temperature 0.6 arated temperature 0.4. That's good. Even reference temperature, I would argue to change that to more. I'll say change reference temperature to 0.9 and aggregator temperature to 0.2. I'm going to presend it. The reason for this is because the reference models or agents, they're like more brainstorming, right? They're uh should be they should be more creative, more random. But for the aggregator, we want it to be more consistent and predictable. So, we want a lower temperature. Temperature is kind of like the randomness. Not really, but that's the easiest way to explain it. Okay. So, it's saying that it's ready basically, but the API key is commented out. So, I'm going to say make sure the API key is not commented out and run Hermes again on the VPS and send another test prompt. All right, so it started Hermes and testing it again. Here at the top, you can see mixture of agents default and aggregator is Opus. And here it is. This is mixture of agents in action. Reference one of four. We have GLM 5.2. We have reference 204 5.5 GB 5.5. Third reference agent is Kimik K2.7 code. And the fourth one is open 4.8 and it's running. This is a mixture of agents in action. There is some error. But the beauty is clo on the left can see it can see everything that's happening right. So what is the thing? M O test successful blah blah blah. One flag non fatal warning. Okay. a fix. Yes, set that. So, it can like proactively uh figure out issue that the auxiliary model is not set up. But it should probably set that to set 4.6 on open router just so we have a more diverse set of models. You know, this is one of these rare cases in the universe where actually more diversity is good. Um I'm not going to go deeper on that, but basically I can just have cloth code figure it out. And next is to test the mixture of agents on something practical because again we get the best models from all of these different AI companies and we can mix them inside of agent to get the best output possible. So for example on back end debugging you know GLM 5.2 and GPD 5.5 are really good on front end open 4.8 it is better than these models, right? So you as a human, you shouldn't be able to thinking like what's the best model for this use case. Just use MOA with a nice preset and you will have the best AI for that use case. But again, it is not cheap because you're running multiple models for all these prompts, right? So just keep that in mind. In fact, we can check open router. Let me reload this. So so far it spent 0.2. It's not that bad. But once we run a bit more impressive use case in a second, this will probably go up. All right. So this is crazy. The cloud code is overthinking it like crazy. So I'm going to just escape it out. And I I don't have patience for this. I'm going to switch to PI agent here, which I'm running GLM 5.2 fast through Versal AI gateway, which is going to be much faster. I'm going to say um pwd. Boom. Check the other pane in this CMAX workspace and let me know if you see it. It should be sshed into my VPS. So as you can see, this is way faster. The inference is like 200 tokens per second. Uh the default cloud code plan without opus fast is very slow. So say okay run Hermes on it and send a test prompt. Next investigate if the auxiliary model inside of Hermes agent has been fixed or if it still has issues. I'm going to presend it. And this is the beauty. You can just swap out to a different agent powered by a different model powered by a different provider. Right? So here the mixture of agents on the Hermes is powered by open router but my PI agent is powered by Versalia gateway using GLM 5.2 fast which actually has much faster inference than any of the endpoints on open router. So I don't know what vers did there but they cook with this one. So okay this is working and you can see the mixture of agents is the default model inside of Hermes now. So it's really good. So let's give it a proper test something harder to build a full stack app and to deploy it and look this is the beauty of having the correct skills. You can see that on the left PI agent is running sleep 6, sleep 10, sleep 12. It's monitoring the Hermes agent and it's pausing anytime it's isn't finished. Right? So, it's not wasting tokens. It's going to sleep and it's holding off and it's beautifully managing the agent. It's observing it. It's monitoring it, making sure it does its thing well. If it got stuck at any point or if it crashed, it would restart it, right? And this is because I set up the skills properly. So, again, all of these skills are available completely for free. Just go to my Twitter. It's in the Twitter bio. It's going to be linked below the video. All right. So, I killed it with Ctrl C. And now I'm going to start a new Hermes instance. Refresh. I'm going to say, I just killed the Hermes agent. I started a new instance. Let me know if you see it. Answer in short. Do not do anything. Yeah, it sees it. Okay. GLM 5.2 is really fast. It's The only downside is that it cannot see things because it doesn't have vision. But this model is really, really good. Right now, it's the best open source model in the world. I have really high hopes for GLM 5.3. Next, I'm going to say tell Hermes to build the following demo, but be very clear about having it built fast and to not overthink and to focus on impressive and quality 3D graphics and visuals. Boom. And then I'm going to do shift enter because inside of Cmax, you can do shift enter. There's a lot of convenience instead of using CMAX, right? Rather than the default terminal. So then I'm going to just paste in project and I have this uh project description. Again, this is going to be the second link below the video right here. Everything I'm showing in this video is in the second link below. Go ahead and grab it right now. It's completely free. Uh back to CMAX here. So build fast do not overink best action and it wrote this temporary file and now it's going to send this as a prompt. So what it's doing is it's creating this file /temp uh herbuilt task and it's going to give it that markdown file in the prompt. Oh, there it is. Boom. And yeah, these agents can write better prompts than me and you as well. So like if you really express yourself properly, use the agent to orchestrate other agents and you'll get better results than you doing it yourself. And again, we have the sleep timeouts. So PI agent is going to monitor this Hermes and making sure it completes this goal. If it's not finished, it's just going to go to sleep and, you know, not waste tokens in the meantime. Check on it. Send me a short update. You know, send me a quick status. What's happening? Hermes is computing. First MOA round still running. blah blah blah and it goes to sleep for a bit longer this time 20 seconds because it realizes that this is going to take a bit longer. Very nice process. Uh again, setting up the right systems for yourself and the right workflows. Essential. It's really priceless in the AI agent era. Let's check the spend so far. $1.8. Not that bad. Could be worse. Could be worse. Let's see how this 3D game is going to how much it's going to cost. Let me zoom this out a bit. So, we see this is how it works, right? You can see the reference agents. I have four, but you don't need four. You could just start with two. That's the minimum amount where this makes sense, right? But I like these four. These are like the four best coding models right now. Basically, you know, two of them are open source. Kimi and GLM, which I would highly advise you guys move more and more of your token spent to open source models. Do not send all of your data to OpenAI Enthropic. These companies do not have your best interest in mind. They distilled all of the web's knowledge, all of humanity's knowledge, art, code, literature, everything, and they created for their own closer models using to generate disgusting amounts of profits. And until then, the social contract was that, okay, they're going to do this. They're going to scrape everything, all of our knowledge, extract data from our brains, and give us this super powerful entity, this very powerful model that can make us more productive, save us time. And most people, you know, myself included, were fine with that. But now they're not even giving us the models. They just took all of that data, all of that priceless training data, and they just creating closest models and keeping it for themselves and for the government. So, it's really tragic, guys. Please learn how to sell those AI models, start using more models locally, start using more open source models. Right? The biggest misconception about models like GLM 5.2 and Kim K2.7 is that number one, they're bad. It's not true at all. These models are really really good. They've caught up to the cutting edge and I wouldn't be surprised if in a month or two GLM 5.3 or Kim free are better than the openly available, you know, publicly available closed models. But the second biggest misconception is that just because the models were trained in China, they have some secret back and some some, you know, hack that allows the Chinese to steal your data. Guys, these are open models. They're open weights models. You can spin it up in a you know 10 20 minutes a GPU cluster with like four B300s and host it in Europe in Australia in USA in Brazil anywhere like anywhere where there's a set of GPUs you can host an open weights model. So start allocating more and more of your resources to these models. As you can see right here in PI agent I'm running GLM 5.2 fast and this is way more affordable than Opus 4.8. So if I was running Opus 4.8 same agent by agent would cost me a lot more money, right? But since it's GLM, I can run it even in fast mode and it doesn't cost nearly as much as Opus or Open models, which are very expensive, the most expensive models out there, and they have disgusting profit margins, right? In fact, I'm going to let you in on a little secret. These subscriptions, a lot of people think they're subsidized, right? whether it's CH GBD Pro or Cloud Max, they people think, "Oh, I'm paying only $200 and they're giving me $2,000 or $4,000 of usage." Actually, that's not really the case. The case is that they just have disgusting profit margins on the API. So, when you look at these Chinese companies, the reason why Deep Seek and Kim and GLM and Miniax are so much cheaper than Anthropic models or OpenAI models is because they don't have these disgusting profit margins, right? like probably over 90% on API usage. So, Enthropic is absolutely printing on all of these enterprise companies that are using cloud models. They have disgustingly high profit margins and that's why the subscriptions can be lower. It's like a lead magnet. They throw you a little bone and they want to make sure that you build your whole business, your whole life on top of their closed models, right? On top of cloth, clothed co-work now clothe instead of Slack. Guys, you need to resist this. Do not build your business and life on top of closed models. You'll absolutely regret this in two to three years where access to AI will be more essential than access to water because AI is going to be super intelligent. You will wish you will wish that you are self-hosting your AI and that nobody has control over your AI models. If you give this power to OpenAI on anopropic, you're going to regret it massively. All right, so let's get practical. Let me explain a bit more about how this mixture of agents is working inside of Hermes. Right? So these four reference agents are giving their opinion. And basically they're giving suggestions to the aggregator but ultimately the aggregator decides what is the best approach. Right? So here we have state assessment uh current state is good. Kim cap 7 is say project skeleton blah blah blah step zero is generally verified. So each of these agents is verifying all of the stages and the aggregator is kind of deciding. It's almost like a court you know like 17th century French court where you have the king and some advisers. So in this case opus 4.8 here is the king. It's the aggregator and he has a bunch of advisers, right? One of them is GLM GB. Second one is 5.5. Third one is Kimmy K2.7 and fourth one is a different instance of Opus with different temperature. And each of them are like saying their opinion about different stages of this project. But ultimately the aggregator decides like yeah for this stage Kim is right for that stage GBD 5.5 has the best opinion. So it doesn't matter what matters is who has the best idea, right? It's a complete meritocracy. All right. So what PI agent is doing on the left is monitoring and it's sending me short updates. You can see that the 3D game is being written. And this is the reason I I build the skill like this is because it allows me to understand it quickly what's happening. Right? If I had to like go here and look at the code like okay there's a renderer you can see that case there's a constant of colors like it would take me a lot longer to understand what's happening. So, it's not all about the output tokens, right? If I speak, you know, speech to text, that's the fastest way for me to output tokens, but it's also about input tokens. A lot of people don't realize this. You need to optimize your workflows so you get the right info in front of you as fast as possible. In fact, that's one of the biggest challenges with complex large scale systems, right? If someone put you in charge of a country, one of the biggest problems you would have is not making the decisions. Like a lot of these decisions are obvious, right? Oh, punish the criminals, right? Do not import unproductive people. like most of us would do a better job managing the European Union than the current EU leaders. The bigger issue is getting the right info in front of you, right? Making sure you're getting truthful information from all the different sources compressed enough that it doesn't overwhelm you, but not compressed too much so you lose important nuance, right? That is the same with managing AI agents. So here when I have my PI agent polling the VPS and polling her mess, it gives me concise updates every 30 to 40 seconds and I can quickly look at it and see like okay this one's proceeding well. Oh no it's going the right direction wrong direction. If I didn't have that imagine you know this left side doesn't work and I just have to it would take me a lot longer to understand what's happening and that would mean my agentic workflow is less optimal. Okay, look at this. This is actually beautiful. The managing PI agent steered it right. Look at this. Look at this. So it was like status still processing frozen at 60k no new tool cost for 3 minutes. Okay. Aggregator appears stuck on a single run. So then it waited and then it sent a steering thing right. So it sent a steering prompt status check. If np is finished type check any errors keep moving toward deploy. So the pi agent powered by GLM 5.2 is doing its job. It figured out that Hermes got stuck for some reason 3 minutes without any updates and it sent a steering prompt, right? And because everything is running inside of Cmax, which is a terminal multiplexer, it can easily access the other Cmax pane, right? So the left side here, the buy agent sees the right side, the Hermes. It knows what's happening. It can also easily spin up new PES, right? So it can launch a new one here, new one here, new one here. Easily it can do that itself because I have a CMAX skill. In fact, I can show you. Let's say uh read let me just do pi read the cmax skill boom skill cmax is one of the skills in my repo and it has the key rules noted how to use the workspace ID how to work with the pes stuff like that right so all of the things that I learned when I first set this up and when I first had agents like pi managing other agents inside of cmax and launching different terminals and different cm panes and workspaces all those lessons I put that into skills So now anytime I mention a cmax py agent or claw code or codex whatever they know to learn to read that skill and the skill has all of the info on how to use cmax effectively. So everything I'm doing here I don't have to mention for it to go to sleep and keep polling the other agent. It knows how to do that right because of my skills which again you can get from my Twitter bio David Andre one. It's going to be linked below the video. Just get this. And while you're at it make sure to follow me on Twitter for the most unfiltered hot AI takes. Let's see what's happening. Uh, status check. Okay. Boom. Maybe the open router API is down for open 4.8. This model had so many outages lately. It's kind of crazy. So, that's definitely possible. All right. So, it's finished. It took a lot longer than I expected. I literally just went on a full walk outside while I let this finish. And uh, yeah, it took like almost 20 minutes, but it should be done. Let's see. Her is asking question. Approve verification. Blah blah blah. I say, "Explain in plain English to me whether the game is done and whether we can play it and test it or not yet." Answer in short. Okay, the upload is already succeeded. Let's copy this. It should be done. New browser. Okay, it's fully deployed, guys. This is pretty crazy, right? So, not only did it build the game, it also deployed it. I didn't do any of this. It built the app itself and it deployed itself on a live URL. I didn't set up anything. I didn't give it any API keys or authentication for it to be able to deploy front and back end full stack, nothing like that. It figured it out. Um, this is the power of AI agents. If you're not building with agents and and like if you're still keeping everything locally and you have you don't have any fully deployed apps, what are you doing? This was deployed publicly by the agent, right? It's it's live. It's live. So, we're going to test it. Let's click start. This is a 3D flappy. Okay. 3D Flappy Bird. We need to do space. Okay, I see how it works. Okay, the graphics are a bit over the top, but the logic is there. Okay, boom, boom, boom. All right, I failed. It works. Okay, we could probably turn down the glow. It's like a pretty extreme glow. And maybe the space should be jumping a bit higher so that if there's a bigger gap, I can hit it. Okay, the hitboxes are slightly larger than they should be. But it's good. It works. And it's fully deployed on the web. literally does the main thing. It's a public website, right? I can open an anonymous window, put it in here and it will work the same and it was all done by Hermes agent with the mixture of agents feature. So yeah, obviously this is not for quick tasks. Let's check the spend reload. $20. Yeah, not the cheapest. Definitely not the cheapest. But this is how you get frontier intelligence. This is how you go beyond the intelligence of Opus or GBD 5.5 by using mixture of agents. And now it's available inside of her as agent. And now you know the steps of how to run it and you can use it on your hardest problems, most valuable problems in your life to get close to fable level intelligence or GBD5.6, which again are not available to the public. And again, if you don't have a VPS, click the first link below the video and set up your first VPS on Hostinger today. It's super easy. It's not only one of the most affordable options out there for VPS hosting, but it's so damn easy. Like, you've seen it in this video. This setup is effortless. And especially if you use CMAX with cloud code or PI agent to help you set up the Hermes and manage it. Hostinger makes all of the things around VPSs so easy. That's why I use it for all of my VPSs. Again, click the first link video, get started, use the code David to save additional 10%. And with that being said, thank you guys for watching and I wish you a wonderful productive week. See you.