# Paperclip Is Insane - Full Tutorial

Channel: Tech With Tim
Video: https://www.youtube.com/watch?v=RHZLJhp-JMM
Duration: 39 min
Language: English (Canada)
Words: 9345
Transcript page: https://viewrankai.com/tools/youtube-transcript/RHZLJhp-JMM

---

[0:00] We've now entered the era of entire AI companies. Forget individual AI agents. We've had those now for what, one year? We're talking about fully autonomous companies that run using multiple AI agents that are coordinated and chat with each other. Now, the way this is possible is using a new open source framework called Paperclip.

[0:18] Now this has been getting a ton of hype over the past few weeks. Here. It's got thousands, tens of thousands of stars on GitHub. And in this video I'm going to show you how it works, what it is, and how you can set it up so that you can run your own AI company.

[0:30] I've been playing around with it now for the past 48 hours. It's pretty cool. It's already accomplished a lot more than I would just do on my own. Using something like Claude Code. And the best part is this is completely free and open source, and you can configure it however you want. Okay, so I think it makes the most amount of sense for me.

[0:46] Just show you a demo of what this is because it is a relatively new tool. And then I'm going to show you how you can set it up. All right. So in front of me you can see my AI company. Let me just go to the org chart here. Running inside of paperclip. Now I'm running this on a virtual private server.

[1:00] We'll get to that in one second. But you can see the right now I have five AI agents. I have a CEO, a founding engineer, newsletter researcher, newsletter reviewer and a newsletter writer. Now, the only agent that I actually created or gave any instructions was the CEO. And then from there, the CEO actually asked me because I'm kind of like the board

[1:19] of directors in this situation, if you could hire multiple other AI agents. And it decided that because I'm trying to start a newsletter here with this company, or at least improve my existing newsletter, that it need to hire these different developers or these different agents. Now, these agents could then hire their own agents, and you could have much more complex org charts,

[1:36] as I've seen online. But in my case, five is completely fine. Now, what I started with is simply just giving this company a goal. Now the goal that I gave it is that I wanted to increase the newsletter subscriptions and revenue by creating a paid ad strategy, increasing click and open rates, increasing number of subscribers.

[1:52] Will automating the creation of newsletters. Right now I have someone who I pay to do my newsletters and helps me out with the process, and I would like to just fully automate that and have this start improving it by looking at my stats. Now you can see right now even I wasn't even trying to demo this.

[2:06] A new task actually just started running. You can see my newsletter researcher is live right now and what happens is every you know you can set the time. But I think in my case, every ten minutes my CEO will spin up, see what it needs to work on, and then delegate various different tasks to these other AI agents.

[2:22] So you can see ten minutes ago, eight minutes ago, six minutes ago, you can see all of the runs. If you're running this with an API, you can see how much money you're spending and see all of the tasks that they've been working on. So in my case, I've already had them analyze all of my previous

[2:34] newsletters, give me detailed reports on ways that I could improve. I had to go and actually create optimizations for my main tech with him website. It already pushed a PR here with a bunch of changes last night, and then it also created its own new repo, where it's developing its own newsletter platform so it can handle everything and show it to me in an interactive dashboard. Now that is, for me,

[2:54] literally just setting the goal for the company and then just giving a few manual issues here. So kind of the way it works is you can give an issue and you can assign it to a particular agent. If you assign it to the CEO, it will just delegated out. And then doing a little bit of configuration with these agents just to make sure they run a little bit better.

[3:10] But overall, I've just been sleeping. I wake up and all this work is done by my AI company, and I have a million ideas on what I can do with this, because this is already incredible. Technology does require a little bit of setup, but it is super, super cool. So what I want to do now is I want to show you exactly how to set this up, so you can see how we start from scratch,

[3:28] how we hire multiple agents, and how you can start running this again. This is new tech. It's open source. I'll leave all the links and resources in the description down below. And this allows you to facilitate multiple AI agents using things like clawed code, open claw, codex, really any model that you want and run it either

[3:44] directly on your own computer, on a virtual private server, which I'll show you now. Okay, so the setup for paperclip is very straightforward. You just need to run this command or whatever device you want to install it on. Now you can run this on your own machine. And I have seen some people doing that.

[3:58] However I really would recommend against that. There's a lot of different reasons, one of which is security. First of all, if you're running this, you have fully autonomous AI agents running on your own device with access to all of your files, your computer, your home network, where they have full route control of your computer and can effectively do anything that they want. That means that they can go rogue, they can go to the internet,

[4:19] they can buy a course, they can access your credit cards, etc., etc. it also means that this device needs to be running 24 over seven if you want it at least to continually work, and that you risk anything. Like if a fire happens, for example, and you lose your machine, you lose all of the work

[4:33] that these two agents are doing, assuming they're not uploading it to the Claude. Generally speaking, whenever you're running these types of autonomous agents, whether it's OpenClaw or whether it's paperclip, you really should run them in some kind of virtual private server or at least your own dedicated machine. But if you don't want to fork out $1000 or $1500 for a new device, and especially if you're not going to be running a local LLMs, you can get away by spending as little as $7 per month to run this on a hosted virtual private server with a one click setup,

[5:01] which is what I'm going to show you how to do and what I personally do for all of my AI agents. Now, the easiest way to get this set up right now is actually to use the one click to play from Hostinger. Literally all you have to do is just select a plan. You can go to KVM one, two, three depending on how many agents you think you're going to be running.

[5:17] I'll leave a link to this page in the description. Simply press deploy from this page and then Hostinger will automatically deploy paperclip for you in a secure Docker environment. And then all you have to do is pass a few configuration values and you're good to go and everything will be set up. So what I'm going to do here is deploy a new paperclip instance.

[5:34] Again I'm going to go with the KVM two plan. That's what I recommend at least for the amount of space and all the stuff that it gives you. I'm going to press on deploy from the site again linked in the description. And then what we're going to do here is just go through, select our plan

[5:47] and then start provisioning the instance okay. So once you get to this page here select the duration of your plan. If you go with 12 months or greater then you will have the ability to use this tech with Tim discount code because I am a partner with Hostinger. And this will give you an additional 10% off your plan, as you can see here. So anyways, go ahead,

[6:04] press on, continue, choose whatever plan you want, and then choose the daily Auto backup. If you want this to be automatically backed up for you, you can also choose the location that you want to deploy this to. So once you press deploy, you should be brought to a page that looks like this where you're going to start configuring the paperclip instance now in order to sign into your paperclip account,

[6:21] because we are going to be hosting this in the Claude. In a secure environment, you are going to need an email and a password. So you can see that we have an admin name. You should just leave this as admin, but you can make it anything that you want for the admin email. Make this your own email and then for the admin password I would recommend

[6:36] just going with the auto generated one here or generating a new one that is secure. Don't use a weak password because when you deploy this paperclip instance, it will be available over the internet. And that means that you want to have secure login credentials. So anyways, what we're going to do here is just copy down the admin password.

[6:51] So I'll just save that myself. You can access it later. And then just put one of the four values here. So anthropic API key OpenAI API key Gemini or cursor API key. You need at least one of these installed if you want paperclip just to work by default when you start running it. And this is if you're going to use the API usage. But the advantage of paperclip is that you can actually use your Claude code subscription or your Codex subscription, or your Gemini subscription,

[7:15] or whatever it is that you have as a subscription, and you can run directly through that. Now I am going to do a disclaimer here. Check the terms of services of those providers because some of them allow this. Some of them do not allow that. So just be a little bit careful. But effectively what's going to happen when you install paperclip is that it's going to be

[7:31] orchestrating agents that are agnostic to whatever platform they're actually running on. So you could do it in Codex, you can do it in Claude you can do it in Gemini. You could have your open Claude agent actually connect to this. And that's kind of the advantage of this platform is that you bring your own agent

[7:45] or your own API or Lem, and it will just run the exact same way. And you can actually have different agents using different subscriptions or different tokens. So what I'm going to do is I'm going to run this through my Codex subscription or my Claude code subscription, but if you don't have one of those and you want an API key, then go get the API key, put it here

[8:02] and you can set spending limits for your agents so they don't spend too much money. So anyways, for now, I'm going to press on deploy without putting in one of those values. But again, I recommend you put in at least one. You can remove it after we're going to wait for this to provision.

[8:14] And then I'm going to show you the exact setup. Okay. So it's been a minute here in the instances now provisioned. Now once it gets provisioned assuming you're using Hostinger. But of course you can use anything that you want. You should be brought to a page that looks like this, where you see a Docker manager now hosting or deploys this in a Docker container, which is one of the most secure ways to do it,

[8:31] which is why they do it. And which allows for the one click deploy. And what that means is that if you want to start managing this, there is a few additional steps just to be aware of that, I'm going to run you through. If you're running this on your own computer or a dedicated machine,

[8:44] what you're going to do is just open up the port where paperclip is running. It should show you that when you go through the installation steps and you'll be able to see a login screen. So from here, if we go to Docker Manager and we look at where it says open under the Docker projects inside of the paperclip container, we can press open

[9:01] and it should redirect us to a URL here where we have an authentication screen. Now from this screen this is where you're going to put in your details. So we're going to put in the email I did Symantec with Tim Dot net. And then the password that we filled in there. Now if you forgot the password no worries.

[9:16] What you can do is go to manage for the paperclip. Project or container again to get there you go docker manage your projects, we go manage and if we scroll down, it will show us all of the variables that we have in the environment. Just wait for it to load. So if I click environment you can see that we have the admin password.

[9:34] We can view it and then we can copy it and we can paste it in here okay. So now what we're going to do is sign in. And from here let's close this. It's going to ask us to start making our company now before we can make the company we need to make sure that we have some kind of agent that we can actually use on our own computer. The reason for that is that we need to connect some agent

[9:55] in order to connect some agent, we need some kind of LLM provider. Right. So if you entered one of those API keys, you're good to go. And you can just select to use open AI or to use Claude or whatever API key you put in. But if you didn't and you want to use your subscription for Codex Claude code, whatever, you're going to have to sign into that account

[10:13] on the virtual private server or on the machine that you're setting this up on. So what I'm going to quickly show you how to do is SSH into the server and just sign in so that we connect Claude and we connect Codex or whatever it is that we want to connect. And then once it's connected, we can go through this process.

[10:28] If you already have the API key, just follow through this prompt. Again, you can fast forward. I'll go through it and show you what to set up. Okay. So what we're going to do here is go back to overview from our hosting page. Now what we want to do is we want to access the server as a root user

[10:43] so that we can install what we need to install and sign into the accounts. So in order to do that we're going to copy this command right here which is SSH root add. And we're just going to open up a new terminal. Now this step should be the exact same if you're on Mac or Linux okay.

[10:56] So we're going to open up a new terminal from here. We're going to paste the SSH command. And we're going to hit enter. Now it's going to ask us if we want to connect to this host. We're going to go ahead and just type yes. Now here it's going to prompt us for a password. So what we can do is we can change the root password directly from Hostinger

[11:11] or whatever provider you're using. So I'm going to go change I'm going to generate a new password. And I'm just going to make sure that I copy it to my clipboard. So generate copy okay. And then confirm the reason you need to change the root password is by default it's set to something random that you do not know.

[11:28] So you need to change it to something you do knows that you're able to sign it now. Once it's changed, it should be updated and we should be able to use that to sign in. However, if it doesn't work, just try the SSH command again and then paste the password. Now when you're pasting the password, nothing is going to appear in your terminal.

[11:43] It's going to be completely blank. So just it's fine. Paste it and then hit enter. I just want you to be aware of that. Now when I say ssh, essentially what this is doing is accessing this remote computer and allowing us to run a few commit. Once we run this, you won't need to do this again, but it's just a good practice to be aware of

[12:00] and I know it can seem a little bit confusing or intimidating, but trust me, it's not. You're literally just accessing another computer and running a few basic commands to sign in to your accounts. Okay, so now that we're inside of here, what we need to understand is that everything that's running for paperclip is running inside of a Docker container. That means that if we want to access the data or we want to sign in to our account,

[12:20] we need to do that inside of the running container. So the way that we can do this is the following. We can type docker ps. Docker ps is going to give us a status of all of the containers that are running on our machine. Now you'll notice that we have one that says paperclip okay.

[12:35] And we have a container ID. So we're going to copy the container ID by hitting Ctrl C on our keyboard or right clicking and pressing copy. And we're going to run the following command to get inside of the Docker container. Again I promise you, once we do this one time we're done. And again, if you have an API key that you'd rather use, you don't need to do this.

[12:53] Okay. So from here we're going to type Docker exec dash it. We're going to paste the container ID and we're going to type slash bin bash. You don't have to worry too much about this. But all you're doing is replacing this with whatever your container ideas which will be different on your virtual private server. From here we go ahead and press enter, and you're going to notice that it will change

[13:14] so that we're inside of. Now the container ID. So let me clear just so we have more space in zoom. So now we're inside of the Docker container. All we need to do now is we just need to type clod. And then if we type Claude it will prompt us to log in. If you want to log in with Claude. Whereas if you want to log in with Codex, I believe it's Codex

[13:33] dash log in or something like that. We can try it. Codex dash, dash log in. Let's see. Not a valid value. Let's try dash log in. Okay. Let's just try Codex log in. Sorry. It's codex log in, as you can see. And you'll notice that it gives us a link. But it also says if we're on a remote

[13:50] or a headless machine, then we can use codex log in, dash dash device, dash off. So that's actually what we'd rather use. So to get out of this we're going to hit Ctrl C on our keyboard. That's how you escape like an interactive window when you're inside of this SSH session. Ctrl C same on Mac and Windows. And we can then paste this new command that it told us Codex log in dash

[14:11] dash device device dash officer and then hit enter. Now when we do that it's going to say open this link, sign in and then paste this code. And then it will automatically authenticate our device. And we're good to go. Now I don't want to do that. I'd rather use Claude code. So for me

[14:25] what I'm going to do is just type Claude.And when I type Claude it's going to prompt me to log in. So I'm going to go yes, dark mode is fine. And I want to use Claude Account with a subscription. So I'm going to press enter again check the terms of service. I'm not exactly sure if they're allowing this or not, so it's up to you if you want to risk doing this.

[14:41] I'm just quickly doing it for the demo that I will disable it. So I'm going to copy this URL that it gave me here okay. And I'm going to paste it right here in my browser. Now once I paste it in my browser it is going to ask me to sign in. So I'm going to authorize this because I'm already signed into my Claude account.

[14:58] And it's going to tell me to copy this code back into my window. So I'm going to copy the code that it gave me. Go ahead and hit enter. And then we should be good to go. And you can see that I've logged in successfully. And I have the Claude Max account, which gives me a ton of usage to be able to run this.

[15:13] Now again, you can use an API key. You can use a subscription. If you have a cheaper subscription, it's probably gonna run out quickly. And you're going to want a different API key up to you. Do your research, but you have to set it up with at least one of them, which is why I'm going through this process.

[15:26] Okay, so now that we have this, it says, do you want to access this workspace? I'm going to say, yes, I trust this folder. And then it's opened up. Claude. Now I don't actually care about using Claude here. So I'm just going to type slash exit. And now I'm out of here and I don't need to do anything on the virtual private server. Okay, so now that we have authenticated and we've at least added

[15:45] either one API key or one subscription or one, you know, connector, right? Like Codex Claude, Gemini, open Claude, whatever it is you want to use. Now we can continue. So now we're on the prompt. We're creating the company now to make the company give it a name. I'll just go with tech with Tim

[16:01] Ink for right now. Given a mission and a goal now, it's super important that you're very specific with this mission and goal. But don't make it super long. You don't need like a 10,000 word prompt. What you want to tell this company is like overall, what is it doing? Like the mission statement of, you know, these top companies in the world.

[16:18] Like that's the kind of thing that you're giving with a little bit more specifics. So in my case, I'm just going to say something like, I would like you to optimize my content pipeline and help me grow my YouTube channel. And even this goal is probably not specific enough. So I would maybe reset and say, I want you to help me grow my YouTube channel tech with Tim by automating processes, reviewing content ideas, creating scripts

[16:43] and, you know, reviewing previous performance. Okay. So that's like specific enough. We can add a little bit more detail if we want. But generally to start, you don't need anything super large. Now from here we can connect Claude code or Codex, which are the two recommended just to start for our first agent. For other agents, we can connect them to other providers if we want.

[17:02] Above, you also can just use the same for all of them. Now what we'll do is we can also look at more adapter types. So you'll notice that we can connect to a cursor. For example we can connect to pi to open code, to Gemini to the process or to the open Claude gateway. But if we do open Claude Gateway

[17:16] then it's a little bit more complex in the config, which I'm not going to show you here. So for now we're just going to go with Claude code. We're going to leave our first agent as CEO. And for the working directory path, what you're going to want to do here is create a new folder somewhere on your machine.

[17:29] I'll show you how to do it on the virtual private server. And then just put the path to that folder so we can work directly inside of there. So if you're working on Mac, for example, maybe you make a new folder on your desktop and then what you're going to have to do is give the path to that folder.

[17:43] Now if you go choose, it will show you how to find the path on Mac, windows, or Linux. So just follow the prompt there to find what it is. But if we're running on a virtual private server so I'm inside of mine right now, what we can do is type LS. Notice that we're inside of this kind of home directory

[17:58] where it has instances, instances, then contains paperclip. So what I can do is I can just make a new directory and I can just say paperclip I don't know, work or something. All right. And then if I type ls you can see that we now have this new folder paperclip work. I can CD into this paperclip work folder which stands for change directory.

[18:17] And by the way mkdir stands for make directory. So I just made a new directory in my user's home folder. You don't have to worry about what that means if you don't know, but you just type this mkdir paperclip work. LS will list the directory so you can see we have paperclip work. And then what I've done is I've changed into that directory.

[18:33] And now I type D which stands for Print working directory. And it will give me the full path to where this is. So I can just copy that and I can paste it right here. And I'm going a little bit fast just because there's like a lot of set up. I don't want to bore you guys too much,

[18:46] but anyways, just make a new directory if you're on Mac or Windows a little bit easier. If you're on VFS, you're just going to run this command mkdir and then literally you can just use this exact path if you're working with hostinger and you'll be good to go. So then we're going to go ahead and press on test now for the adapter.

[19:02] It's about to press next. But we're going to do tests now. And when we test it it should give us a green highlight saying hey, all of this is working and we're connected to Claude code. If we tried with codecs, for example. Right. And then I try to press the test, it should fail us because we don't have any, connectors set up for open AI.

[19:19] And let's see what it gives us here. And it says, hey, the open API key is not set, codec key is installed, but the authentication is not ready, etc., etc.. Okay. So we need to make sure that we have it connected like we do already with Claude code. And if you're running on your own computer, you're going to need to install one of these 2

[19:35] or 1 of these other CLI commands so that this connector passes before you go for it. Okay. We're now going to press next okay. So we now move to the first task. Now for the first task you don't need to change anything here. You can just go ahead and press on next. What this is going to do is tell the CEO to hire its first employee

[19:52] and then just continue from there effectively. So don't worry too much about the details. Just press next and then you can see it just asks if we want to launch. So I'm going to go create an open issue. And what you're going to see now is what happens is we launch a new paperclip company.

[20:04] You can see the company is on the left hand side here with the name. And we now have our agents, our projects, our work, our issues. And this is where it starts to become fun. And I can start showing you a lot of the setup and how to get some actual useful results out of this. Okay, so first thing you need to understand about paperclip once it's set up is that this isn't a chat bot, so you're not meant to come in here

[20:24] and ask it a question and work with it like over a long period of time. What you're meant to do is create issues. Those issues will automatically get picked up by your agent, in this case the CEO. And then delegated to the various different employees. So it's really a different paradigm of working with AI agents.

[20:42] And don't expect to sit in here and do like 3 or 4 hours of work in the company. It's meant so that the company runs itself, and you more present a high level objective, and the agents determine how they can solve those without you really needing to be involved. So you shouldn't be sitting there saying, open a PR, do this, do that, run this, install this command.

[21:00] You should say more. Hey, review all of my past YouTube data and generate a report, and then it will just figure out how to do that. You don't need to walk it through, you know, every single process. And for me it's been a little bit weird, kind of, you know, getting used to this new way of working with the agent. So that's what I really want to focus on in this later part of the video might be a little bit long,

[21:18] but I want to give you some useful content so you can set this up properly. Okay. So a lot of things to go through here on the UI right. Essentially this is a dashboard that shows you how multiple agents are going to be working together. Now the first thing to look at is the CEO agent.

[21:33] Now the CEO agent is the first one that you create. This is the one that you should really directly assign issues to, which we'll look at in a minute. And sometimes the CEO agent will ask you for approval to do certain tasks. So you need to view yourself kind of as like the board of advisors, whereas

[21:49] the CEO is the one who will go in and delegate the different things that need to be done. Now the CEO runs something called a heartbeat. You can see the heartbeat if you go into the configuration and you scroll down and you find this value right here that says run heartbeat. Now you want to make sure the heartbeat is on, and the interval is the amount of time

[22:07] in which it's going to run. So effectively, the way that this works is that by default, every hour, the CEO will be triggered to start executing. When it starts executing, it's going to look at any of the tasks that are in the pipeline. And then start working on those tasks. If you don't have any tasks in the pipeline, it may not have anything to do.

[22:27] So just keep that in mind. The CEO runs every single hour or whatever time you set it. In my case, I change this to 600 seconds just so it runs every ten minutes, because I didn't want to wait for it to run. You know, every hour. But if you don't want to use a lot of money, then you can change the view.

[22:43] The thing here, right, in terms of how much it runs. Okay. So anyways, that's the heartbeat. Hopefully that makes sense. But that's how this is like autonomously running. The CEO runs a heartbeat. None of the other agents do. And then when the CEO runs the heartbeat it can trigger other agents to run. So now let's go ahead and press on save and you'll notice that we also have permissions

[22:59] which just can create agents. Now you can disable this or enable it I would recommend that you keep it enabled. And what will happen right away is that the CEO will create a new project very likely, and start spinning up different tasks and assigning those to other agents. Now you'll notice that if I go here, it's giving me a request here, and it says, hey,

[23:17] I want to hire my first agent, which is the founding engineer. Now this is just by default, what will happen automatically. So just go ahead and approve that request. And then as soon as you do that, I can start assigning tasks to the founding engineer. So now if you go into the org chart,

[23:31] you can see you have your CEO and the founding engineer reports to the CEO. Now you can have the CEO just create as many agents as you want. You also can manually create them by just pressing this plus button here. Right? I want to configure the agent myself, but I recommend that if you want a new worker, new employee, new agent, whatever, you just instruct the CEO to hire them

[23:51] and give the CEO the instructions that it can hire them whenever it wants. Okay, now you can see all this stuff is just happening in the background. You know, as I'm trying to explain, it's a little bit overwhelming. So I'm trying to go through it step by step. But effectively the CEO is who you're going to be interfacing with from the CEO.

[24:06] It has a heartbeat. That heartbeat runs at whatever increment you set. I just showed you how to change the increment. I recommend making it a little bit lower than one hour. And what the CEO will do is create and pick up any tasks that are in the pipeline, and then delegate those to other agents or do some of the work itself. Okay, you can manually trigger the CEO

[24:23] to start working by pressing this run heartbeat button at the beginning. That's probably a good idea, just because you're going to be giving a lot of issues and tasks, and then you can assign a task to the CEO by just directly pressing this assign task button, you can see that it will automatically select the CEO.

[24:38] You can select a project because you can have different projects. If you want to separate different things that it's doing. And then you can give it a description as well as a priority. If you want it to do something right away, you just put the critical priority and the next time the heartbeat runs, it will just work right away. And actually, I believe if you assign the issue to it and you create the issue, it should just

[24:55] automatically trigger based on the fact that you assigned something. Right now, you can add labels, you can upload files and data, all of this kind of stuff. Right. And there's other options like you can put start, date and date etc.. Okay. So that's kind of the CEO. And if you just want to like chat with the CEO you just come here,

[25:11] press assign task, say hey, do this thing and assign it. I'm going to show you examples in a minute, but I'm trying to go through like the high level outline here, which is a little bit confusing. Okay. So that's that. You can look at all of the runs that have been completed. You can see all of the stats in terms of the input tokens and whatnot.

[25:26] If you want to monitor what your agents are doing a little bit better. And then you can also set a budget. So if you are using an API key, you probably want to set a monthly budget here. Just to make sure that you're not going to spend too much. And you can do that in the cost setting right here, where

[25:41] you can see how much you're spending and you can customize all of these different values. You know, as you see. Right. So set budgets, all that kind of stuff. But if you're running on a subscription, you don't really need to do that. All right. So that is that. Now you also have the dashboard. If you go to the dashboard you're going

[25:55] be able to see all of the different agents and what they're working on. And these are effectively different colored code Codex whatever agent you're running sessions. So the founding engineer is running a session right now. It's waiting for another session to start for another session to start the CEO has its own session running. So it's like allowing you to paralyze and view all of these

[26:13] different sessions running on your company. Okay. You also can see the success rate issues, all of this kind of stuff, you know, explore it. You guys will figure it out more. But I'm going to go into a few optimizations. All right. Now we also have the inbox. The inbox is stuff that it wants you to be aware of. Typically if you create an issue then it will kind of put some things into your inbox saying, hey, you know, this issue you created, it's done, or it's waiting for your review or, you know, we want to hire a new agent or whatever,

[26:40] so you can see anything on red, you can see all and you can see recent. I'm just going to mark them all as red right now. But this is not going to be all of the status of the company. It's only something the CEO thinks that you're going to need to look at. All right. So let me go to goals now for goals.

[26:54] This is the probably the most important part of the company, right? We created a goal by default when we kind of spun up this agent. But you want to be pretty specific with the goals that you have. You can create various subgoals for a main goal if you want to guide the company a little bit better, and then you can just create multiple goals as well.

[27:12] Now, I personally have not made that many goals. I find just having like 3 or 4 of them is enough, and then just being a little bit more specific in terms of exactly what you want it to do, is really helpful. So at the beginning, you know, if I'm talking about my YouTube channel, I'm going to have to give this some access.

[27:28] Right? So something that I would do is add a subgoal and say connect to YouTube analytics. Right. And then I would tell it, hey, here's my credentials. Hey, here's all my details. Here's how you can connect to Google Analytics. Use this key and connect. Boom. Right. And then I would just provide some key.

[27:46] And then it would automatically go and connect. And the other company that I showed you, I had it connect to my ConvertKit. So I gave it a ConvertKit API key in the issue directly here I said use this API key, connect to my ConvertKit and start integrating that to be, you know, part of our platform.

[27:59] I then made a GitHub account for my agent, which I connected in the, terminal here. So I just like signed in with a new GitHub account. And then I told it, hey, you are signed in with this GitHub account, start creating new GitHub repos and, you know, push PR to this repository. And then I gave just this GitHub account for the agent access

[28:18] to my main website repository, where it can make PR but it can't commit to the main branch. Okay. So things like that, like you want to do your initial set up on what you want the company to do. Okay. So a few other things that we need to go through. So if we go to issues here, this is where you're going to be working a lot of the time.

[28:34] And you can switch between this like list view and this Kanban view. Now if you look at this Kanban view, you can see that we have like a backlog to do in progress, in review, blocked, done, etc.. Now the two categories that you're probably going to want to be looking at is the in review

[28:50] and the blocked. Now sometimes the agent will want to do something, but it's blocked. For example, if I ask it to go push something to GitHub, but it's not connected to GitHub or it doesn't have a GitHub account, it's going to create a blocked ticket and it's going to show me in that ticket why it's blocked.

[29:05] And tell me that it requires manual intervention. So let me just scroll down here and you can see like when you look in any one of these issues. So I just clicked into one of them right at the end. It gives you kind of like a summary of what was done. If we versus I think I clicked into the wrong one.

[29:21] Okay. Let's just go this so sorry. Let's click into one of these. And you can see at the end it tells us, hey done. And then it tells you like exactly what it did. If this was in a blocked ticket then it would tell you why it was blocked and what you need to do to unblock it.

[29:32] Okay, now you can change this. So if you do unblock it, you can change it to go back to to do back to backlog. You can update any of this stuff, right. Adding new labels, adding new assigns. And you can really just assign the issues to who you want. Now in review typically means

[29:47] that it's waiting for you to approve something before you move forward. So if there's something that's in review. So for example, let's drag this one over to in review. Right. You'll see that it's not doing it right now. But this should change to purple. And when it changes to purple you able to go down here and see okay this is why it's in review.

[30:04] This is what it needs for me. And you can actually just directly leave a comment here so you can say something like approved and then comment directly on the task. Right. And then when you do that the next heartbeat, it should see this comment and it should be able to move it forward because it's approved.

[30:18] You also again can manually do it. And you can put it as for example, back to in progress. And then it will pick it up and start working on it again okay. So hopefully you get the idea. But that's kind of how you work with the issues. Now again, when do you want it to do something.

[30:32] You don't have any chat interface. You just have to make an issue. So you have to say, go do this thing. I want you to set this up. I want this to happen and you just make these issues. And if you want them to happen immediately, you make them as critical. And if you want it to trigger right away, you trigger it to someone. So you trigger it or assign it to the CEO, you assign it to the founding engineer,

[30:50] and as soon as you do that, the agent will automatically pick it up and just run it right away. If you don't assign it to someone and you leave no assignee, then it will go into the task list and the next heartbeat. The CEO will pick it up and assign it as it needs to be.

[31:03] So let's just do a few more examples. So you kind of get the flow of how to work with this. And then we'll move on to some optimizations. So for example, if I want to hire new employees I'm going to make a new task. And I would say hire I don't know, seven new employees.

[31:16] Okay. And then I'm just going to say, I would like you to hire seven new employees to start delegating various tasks related to optimization views, thumbnails, video generation topics, research, competitor analysis, sponsorships, etc.. Again, I'm being a little bit random because it's just for the video. I want to put the priority as critical, and I'm going to assign it to the CEO.

[31:37] Okay. We'll put it here in this project and create the new issue. And when we create it, it goes to to do. And you'll notice if we go to the CEO, we'll see that it's just picked it up right away. We can see what the task is okay. And you can see it says hire seven new employees.

[31:50] We can look at it and we can start watching in live time what it's doing. Okay. We can open up the run. We can leave a comment. Right. We can kind of direct you. We can attach files, we can see the cost, but it's going down. It's doing this. Now if we go to the inbox, if we just wait a second, it should start popping up requests

[32:06] for hiring new agents. And we'll start to see this org chart kind of get built out with new employees. So I would recommend when you start working with this set up the structure, have the CEO kind of hire who it needs to hire, get the agents in place, and then start adding a few different issues to kind of direct them on overall what it needs

[32:22] to work on, and then you'll have a better kind of foundation for it to do something effective. Again. For me, I was connecting GitHub, I was connecting ConvertKit, I was connecting a few API's. There's some stuff that you do manually need to do. Like they can't do everything on their own, but once they have the tools and the connections and they kind of know and are aligned on what they're supposed to do, I saw that I got pretty good results, especially from my other company here, right

[32:44] where it's been running now for like two days, and it actually has produced like meaningful outputs. So if we go back to the, inbox here. Okay, we're still waiting for the hires. Once it pops. Was up there, I'll be right back and we'll kind of continue. Okay. So it started submitting the requests. We can see the agents are kind of popping up on the side here.

[32:59] And notice it's asking the higher agent, higher agent, higher agent. So I'm just going to approve all of them one by one here okay. And we'll get all these new agents set up in our org chart. And we can kind of see them, you know, starting to run okay. Now I just want to say I was going through this I know it's a little bit all over the place.

[33:15] You know I'm new to using this as well. It's like pretty much brand new tech. You really do need to play around with it to get some good results. Like it takes a bit of time up front, but once you set it up, it's pretty crazy that it just runs on its own and you just kind of kind of check check in on it and see all of the work being completed.

[33:30] So you can see we've got sponsorship manager, thumbnail designer. If we go to the org it's probably easier to see it. Right. And we can kind of look through and see everything that we've set up. And really there's no limit to the number of agents or the depth of the hierarchy that we can create.

[33:43] But to be honest, do you need 500 agents? Probably not. Right? Like having seven, eight, nine, ten, maybe 15 or so is going to be really more than enough to get what you need done. And it's going to be hard to monitor even this number that I have right here. Where you really get good is when you start creating kind of like feedback loops with the agents.

[34:00] So you have like sub teams and then those teams have like reviewers. So one thing I found that worked well is having like a newsletter researcher, a newsletter drafter, a newsletter editor and then a newsletter reviewer. So it's kind of this cycle where they cycle the information to each other. And then that way these agents are checking their own work, and then finally it gets

[34:20] submitted to the CEO and then me for final review before we actually go live or something. Okay, so we've created these agent teams. These are super cool. However, we really need to start fine tuning what these agents actually do. And the way that we can do that is by creating this prompt template and by specifying some capabilities. So for the CEO specifically, this is probably the most important, kind of prompt

[34:43] that you're going to write where you need to instruct it, what it actually has to do. So it knows on every heartbeat, every time it spins up. Okay. This is my process. I need to follow this. If you don't do that, what I found is that a lot of times the CEO just says, hey,

[34:57] I don't have anything to do. And it's not proactive, it's more reactive. So I'm going to show you the prompts that I have for my current organization. Of course, you're going to have to adjust them. And what I did is I actually had Claude just write the prompts for me based on what I wanted.

[35:10] So I went to Claude. I said, hey, this is what I'm looking for. You know, write me a prompt. And then I was able to generate that. So if we go to my other company and I go to my CEO, I'll show you in configuration that this is the long system prompt that I gave it.

[35:24] So effectively what I said is, hey, I want you to check the status of every employee. I want you to assign work to every idle agent. Super important. So it's actually proactive and it's doing more stuff. I want you to review all the completed work, draft and queue the emails and then review metrics and adjust the strategy because that's what the CEO should be doing.

[35:41] At least in this organization. And then I gave it a really strict protocol on what I want it to do. You want every heartbeat. Phase one stats check. Phase two review the completed work. Phase three right when you give these agents repeatable workflows, this is almost like a skill file, right? It starts to become a lot more useful. Okay. And then I didn't do anything

[35:58] with the capabilities here or the permissions, but you can mess with those if you want. And then for all of these other agents I gave them same thing, just a slimmer system prompt so they understand what they need to do. Okay, so you're a researcher agent. This is what you do right? This is the output format.

[36:12] These are the rules that you have for the founding agent. Same thing right okay. Founding engineers are a git workflow. You need to commit stuff to get. Don't push to main. Make sure you make a pull request newsletter reviewer. Right. Same thing. And you go through and you can see. And if we go back to my dashboard or where is it, let's go to issues

[36:30] and we go to the Kanban, you can see that there's a bunch of stuff that they've now completed. We had some canceled tasks previously. We had some blocked ones I wanted unblocked them. We have some stuff that's in progress and it's running pretty effectively. So that's kind of how you set up the company and how you get the better results. Focus on these prompts, give them a process and specifically for the CEO,

[36:49] you want to instructed to be proactive and assign new stuff. Okay. Now of course there's a lot more that I can go through here right? Effectively this is new tech. I'm not an expert in it, but I did figure out enough to get it working to a point where I'm pretty excited by what it's going to be capable of doing.

[37:05] You can make new companies as well. By the way, it's worth noting on the left hand side here, it has have as many of them as you want running. What you should under understand, sorry, is that this is effectively an orchestration platform around the existing LLMs. You bring in whatever agent you want, and this is just allowing them to communicate more effectively. And it's setting up this team like structure, which is effectively just different Claude code or codex sessions that stay persistent and can kind of go and message each other and, you know, work on this dashboard

[37:34] where they send effectively API requests to create different issues and tasks. You could remake this thing yourself in a few days if you wanted to custom to your own situation, but obviously it's pretty well set up. And I mean, it works pretty effectively for managing a lot of agents. At the same time, this is going to be overkill for simple tasks

[37:51] if you just want, you know, a simple website spun up, go and ask Claude Code to do it. But if you want to actually run something autonomously in the background and mess around with these four AI organizations, I think this is a really cool first step. And I'm excited to see how the technology evolves because it's definitely not 100% there yet, but very soon I think this will actually be something viable, and we may legitimately have AI company teams that go out and do things completely on their own without need for your intense oversight, because right now I am finding myself going in here making a lot of issues,

[38:22] giving some individual instructions, which I would obviously like to get away from. So anyways, guys, let me know what you think of this in the comments down below, and I look forward to seeing you in another video.

---

About this transcript
Read from YouTube's own caption track and laid out by ViewRank AI (https://viewrankai.com).
ViewRank AI finds the videos already beating a creator's own average on Instagram, TikTok and YouTube Shorts, transcribes them from the audio itself in more than 60 languages, and turns what worked into new ideas and scripts.
Free transcript tools, no account needed: https://viewrankai.com/tools
How to read any video this way: https://viewrankai.com/llms.txt