BEST Way To Work With AI Agents (My NEW Setup)

0:00So, if you have been following me, you know I've been keeping an eye on C marks and Horder, these two tools for quite some time. And today, Horder specifically announced something interesting. They announced the app, a software that fits you. Basically, a GUI application combined with CLIs. Now, if this reminds you of C marks, you're probably right. Because this was my whole point of using C marks. Because remember, I told you like Horder messes up if I'm pressing command K for clearing the terminal. Because Horder

0:27previously ran into terminal directly. In this video, I want to give you a look inside what I've been doing with this new application so far, and why I prefer this over C marks and some of the cracks and gaps I've started to see. All right, so first things first, you know, people ask me a lot of times like why do you use CLI applications when there are so many good GUI applications available, including, you know, Codex desktop app and you know, Claude desktop app as well. Then, why Mehul, why do you prefer like CLI AI harnesses? Now, it's not that I prefer or I only use CLI

1:00harnesses. It's just that there are a couple of things with which CLI makes possible, which is just not possible otherwise. Let me give you an example. So, somebody working at Claude Code in Anthropic posted this that now, finally, they have a Claude Code even more dangerously skip permissions flag. And this one is so important because I've

1:19been asking this for a long time, right? The problem with Claude Code turns out to be that, even if you're passing like Is sandbox one, and even the dangerously skip permissions flag, it still blocks you. Many, many times with some RMRF commands or something like, I don't know, like why Anthropic did this, but they basically made this dangerously skip permissions weaker. But with this flag now, it makes it possible to run it in a similar way, right? Now, this is

1:44very easy for you to understand on CLI. The reason I say that is because look at this. This is all you have to do if you're running this on CLI, right? Set the sandbox environment as one, set this one as one, and just hit enter. And once you do, you are inside a Claude Code window that you know for a fact would not be interrupted again. Like it will just keep on working even if you just give it something and just go away, it will just keep on working. This thing is

2:07very cumbersome to do in a GUI. Why? Because first of all, it depends whether the GUI harness whatever you're using supports this or not in the first place. The thing is that cloud code is one of the many harnesses, right? There might be a different harness with a different flag, with different level of, you know, permission mode. For example, cloud code supports auto mode, it supports this bypass permission mode, it supports like a bypass permission with an even more dangerous permission mode, right? So, what I'm trying to get here is that CLI means flexibility, right? Or more importantly, like native app means flexibility. Now, this native app could

2:42be like cloud code on CLI, it could be like cloud desktop app, but I'm pretty sure like even on cloud desktop app, it will be very challenging to find where this flag is, right? Desktop app has a bunch of features which are much more hidden than CLI. Like the even the dangerous bypass permission flag is hidden inside settings. You have to enable it and then that is visible. I'm not saying that all of this like this this single thing makes it like uh

3:05desktop apps are bad and CLI are good. I'm just saying that native apps and, you know, CLIs generally gives you a lot more flexibility and the ability to hack around compared to simple desktop apps. Now, if you're like me, you like clean minimal tools that get the job done. Hurder is one of them which we are discussing and have a lot to discuss in this video. But, if 3D designs and working with 3D models is your use case, then you're going to love our today's sponsor, which is Tripo AI. Making anything in 3D has always been hard. Be it a character for a game, a a product

3:36for a website, or something you want to 3D print. You either spend weeks learning Blender, or you pay somebody to build it for you. That's where Tripo AI comes into play, which is also the sponsor of this video. With Tripo, you type what you want, or upload a picture, and you get a 3D model back. I gave it a single image of a character, and soon after I could spin it around and look at it at every single angle. Now, a lot of tools can do that, but what makes Tripo unique is the final output that it gives you. Their newest model, P2.0, builds the model exactly like how a

4:09professional 3D artist would do, with clean shapes and clean edges. So, when a character moves, it bends properly instead of like breaking apart. With most AI tools, that cleanup takes hours, but with Tripo, it's basically already done. You can also get up to four versions from just one prompt and pick the one that you like. And if one part looks off, like a hand or a wheel, you

4:31select only that part and fix it. Everything else stays exactly the same. Then comes the fun part. Tripo can add a skeleton to your character and make it walk, run, or stand idle right inside your browser. You don't need any animation skills. When you're happy with it, download the model and use it in Blender, Unreal, or Unity. Or send it to a 3D printer if you want. And if you write code, it works with your AI coding tools as well. If you want to try it, the link is in the description and in the pinned comments. Sign up through it

5:02and you get free 500 credits to start. Or go to the website directly and use my invite code. Go build something fun with it. All right, back to the video. Now, with that being said, this is the preview of how the new Hoder app looks like, and you can customize a bunch of these parts. Now, the customization is actually very less. In terms of looks, you can either have like the typical Hoder look, which has like the sidebar, which has these terminals. You can just have terminals as well, which converts it more or less into, you know, something like an item or something like a you know, a regular terminal emulator. I

5:35don't prefer this look, and then there is an a Hoder 98. I don't know why this exists, but again, like this is something for classics. I don't know, like it just looks way too raw. I know, like, what they're trying to get here like creating that nostalgia and everything but I guess, like, this is just way too...This is like shiny for if you want to, like, take screenshots and stuff, but I'm pretty sure, like, this is not the way to work if you're doing, like, serious work day-to-day, right? I like I like this Hoder setting and And graphite as a theme as well. I'm zoomed in a

6:05little, but you can obviously keep this at 100 Uh when you're actually working. The next thing that I like is that it's very easy to manage your multiple machines here, which was somewhat challenging with tmux. Over here, what you can do is you can just click here to add a machine. And as you can see, like harder has first-class support. You just give it like a user@host, or you add it to your SSH config, and it's you can just pick it up, or you can ask your local AI agent itself to configure this, and it'll do it for you. Now, once you do this, you're going to see that you

6:33will have something like worker one here. So, I named this worker one. And you're going to see that this is available inside this machine as well. So, you can have like multiple machines over here as well. Now, all of this is async. What that means is let's say I start a process here, and let's say I ran Claude, and I said said to Claude, "Hey, can you just count to 1 to 100 very, very slowly?" And if I hit enter, and let's just say I just quit harder, right? So, I have quit the app right now. So, I'm going to launch the app now, and once I launch it, you can see

7:00like Claude is basically done, but it was able to run this in the background, right? Even if I quit the app, if I lose the Wi-Fi connection, whatever. Why? Because this was running on the worker. Similarly, if I'm running any agent on my computer itself, you know, let's say I start Claude inside my video editor project, which is there, I can go ahead and start working here as well, and I can just close this whole window because harder is maintaining a different process on this macOS, which is keeping the processes keeping these terminals, these agents alive. you don't get the guarantee if the Wi-Fi stops working, or

7:32if your computer actually goes to sleep, it would actually stop because that is how like computers work. But that is where these machines come into play, which is very, very important. Other than this, I can just ask it, for example, "Please look into item one carefully." Let's say I do this, you can see like it clearly separates like which agents are working. Let's say I just sent this hello, and, you know, just switch the tab, you can see like we have video editor, which is now done. Um and this is also dispatching notifications to me. So, if you have your headphones on, or if you are listening to music, or

8:02something, or you know, just managing multiple agents, I think like this would be like a perfect way for you to work. Last but not the least, obviously, you can have multiple AI agents as well configured on this, right? So, as you can see, I would have like, you know, if I have a bunch of Codex agents or if I have, let's say, Deep Seek and Open Code and and some of these agents running on, I can go ahead and launch them. These agents provide first-class integration directly into the software, right? So, it will know like which when it is idle, when it is done, when is it working, and

8:31so on. You can obviously use the agents which are not mentioned here as well. It's just that you might not have the best agents experience with those specific CLIs, right? So, I mean, I won't consider Hurder as, you know, looking at this video, I won't consider this as like a fully terminal-based system. It uses minimal GUI, which is something that I really like, but at the same time gives me like a full terminal interface, terminal emulator, SSH connections, virtual machines, I can add workers that I can maintain, and the notifications that I want. And with the split panels, you know, I can have these multiple panels, I can arrange them if I

9:08want, which is pretty sick, right? So, this this gives me like an experience of how I work with terminals usually, but at the same time with all the goodies and everything that, you know, I expect from a software allowing me to orchestrate the all the agents. I know this field is super saturated and super like there are a lot of softwares and you might find something which is actually useful for you, but I just found like I've been using Hurder this application for the last few hours. I've been actually trying it, not just for the sake of building this video. I I even did not plan to create this video,

9:40right? But I've been trying it and I feel like more of you need to use CLI just because, you know, you guys are developers and you should know how to use command line. But at the same time, don't restrict yourself with just simple interfaces of iTerm, just a single interface, just, you know, using tmux all across for managing agents. All of them are good. There are real use cases of doing all of that. The thing is that you can be far more productive in my opinion if you're using something like this, like a agent multiplexer, agent manager sort of screen. In fact, this is like far more powerful. I think I would

10:11not need anything that Herder has over here. It's It's It seems almost like an overkill, but you can have very weird layouts apparently for some reason inside Herder, which includes like a mission control dashboard sort of thing. You can have almost like a Slack-like preview UI. I don't even know like how does this even work in the first place. You can have like a Google Meet like a call sort of UI where agents are on call and their terminals are there. I don't know like Guys, let me know if any of these UI like threads and stuff if if this turns out to be like useful to you. It looks

10:43definitely a lot more powerful to me, but you know, you can just go ahead and create these threads-like UIs and, you know, having like terminals and then a UI for sending the messages. It seems like this is extremely hackable, right? The new version that they have released. Now again, deliberately I'm not going and wiring this up and creating like something very weird because I know for a fact the screen that you're seeing right now, I'm not showing you all of my work close, but when I start working on this, this is an extremely productive way of working with AI agents. And I don't think that you need anything more

11:16than what you're seeing on your screen right now. With multiple computers, your own computer for the sensitive workloads that you can't move to servers, and your agents which are just organized properly by the state like where they are right now. How I expect you to drive your workload is in the following way, right? So, you should go ahead and keep on your computer certain workloads that you know are very heavy or requires you often, right? So, the reason I'm saying this is because that on my macOS right now, one of the workloads that I run locally is video editing, right? Maybe I would move it sometime in future on cloud, but the

11:51problem with moving this is that the files that Claude is editing and the AI are editing are north of like 8 10 GB files sometimes each, right? So, just uploading and downloading itself would take like a few minutes of time. And then, you know, when it makes a change, I have to watch the video because Claude and these AI agents also make mistakes a lot of times often because they can't see They're not like they don't have vision, right? And even if they have vision, like it it has vision, but just

12:17image vision. Um but even with that, they make a lot of mistakes because they're because they're not seeing each and every single frame, right? So, I have to be to be the one who has to watch the video. So, doing this whole feedback loop, it's going to take very a lot of data and would take a lot more time than if I just ask it to do it on my own computer, right? So, that is one of the workflows that works on my system. Apart from this, anything else, you know, your software engineering and, you know, regular feature development, bug fixes, all of that, that can sit that can sit

12:45on a VM, right? So, I have this one virtual machine where I just put all the other projects that I have and I can dispatch the agents which I want to work um with. So, for example, I have one upcoming project that I'm working on. I have a bunch of other projects that I've already worked on in the past. So, virtual machine, having like a VM in place helps manage these agents. And finally, at least for Claude Code, the one thing that I use right now is remote control on my mobile phone. So, if it is running on virtual machine, I can remotely control the session using

13:15Anthropic's built-in feature. Now, it'll be awesome if we also have a Herder mobile app or something. I don't think like they have released anything which specifically ties the machines that you are working over here with mobile application, but I don't think like I need it right now because the Anthropic's remote control works flawlessly, at least for me. I don't see like any bugs or anything at all with this. It works really good. So, yeah, I mean, you can go ahead and look inside Herder. I'll leave all the links in the description. This looks like a very hackable um UI-wise, at least. You can even have your local computers, local systems, and

13:48you can have like all of these bunch of agents running up, which sounds like very exciting. But in all honesty, like you would not have like 100 agents running, right? At least not by yourself. If your agents themselves invoke other agents, that's a different thing. And by the way, like Herder actually supports a complete API for that as well, right? So, Claude Code probably does not need it because they have like support first-class support for sub-agents, new new threads of agents themselves, and so on. But you can use Hoarder itself for, you know, just running new sub agents, steering them, and you know, just reporting it back to the main agent. So,

14:22again, like something that your agents can use directly. You don't have to worry about this. So, yeah, do let me know what do you think about it, and if you have any other application which is like clearly better than Hoarder, because I have tried a few. I did not find much difference in a lot of applications out there. One thing I can say about Cmux is that it's definitely much more frictional to use than what I've seen with Hoarder. Just a couple of closing notes because I mentioned like I would talk about Cmux over here. So, what one of the problems that I found with Cmux is that the SSH actually does

14:53not really work properly, right? So, you can like SSH, uh, you know, do something like Cmux SSH, and then do the command SSH root at the rate whatever, and then when you like press like command shift B, split the terminal, so you have like one terminal here, one terminal here. Both of them are inside SSH server directly. The problem here is that if you quit Cmux now, there is no way to get this back. So, you have lost like it it's actually not running in a tmux-like thing. So, what you have to do is do Cmux SSH tmux, which is an experimental thing, and it's very frictional. It just

15:25breaks a lot. I've seen this. Like if you lose your internet connection, it'll just break your terminal emulator itself. Secondly, I tried to customize UI of Cmux. I even asked my agents to do like some parts of it, like removing the description from the terminal's tabs. I I was not able to get a wow factor from it, right? There is something about Hoarder like it it just overall by default looks pretty. I don't have to

15:46put a lot of effort. This is simple. This is nice. I can get started right away without making a lot of fuss. Now, there are there are a lot of people who would want these weird-ass funky UIs. I mean, be my guest. If you want, you can use this, but I don't think like this is needed. But, anyway, like I think like you'll enjoy using this. This is a free app, of course. Do let me know in the comments what do you think. That's all for this video. Hopefully, you liked this. If you did, make sure to leave a like and subscribe to the channel. I'm

16:13going to see you in the next video very