I Built an AI Agent Company (From Scratch)

0:00How do you turn a group of AI agents into a company that actually runs itself? You know, with org charts, roles, goals, budgets, and agents working around the clock 24-7. There's this project called Paperclip. Maybe you've seen it because it just blew up. It got 30 ,000 GitHub stars in the last three weeks. And it's an open source project that's building the orchestration layer for exactly that, hiring a team of AI agents. So I sat down with Dotta. He's the creator, and it's one of the first times he's publicly talked about how to use his product. We broke down exactly how to use Paperclip in practice and what a zero human company could actually look like using Paperclip.

0:43Enjoy the episode. the startup i just podcast it's in time baby dota is here well his ai avatar is here and we're going to be going through paperclip the viral you know what are you calling it yeah so it's a uh agent orchestrator for uh zero human companies so it's a tool that you can use to put in your ideas and uh manage a company of agents. Okay. And by the end of this episode, what are people going to get out of this? Yeah. So by the end of this episode, you will be able to know how to get Paperclip up and running and use it to manage a team of AI agents to help you with your business. Okay.

1:30And you're not going to hold back from people, right? You're going to give us the unfiltered how to get started, how to get the most out of it, and maybe where the future of this product is going, right? Absolutely. Yeah. I'll give you my best tips. Um, you know, I think there's a lot of tricky things when it comes to using AI agents and we'll talk about it. We'll talk about ways to keep them moving. We'll talk about ways to get quality work. Um, I'll give you all my best details. And, and just, just, I have to ask like, why are you an avatar and why are you not showing your face? Oh yeah. Um, so, uh, because before I worked on paperclip, I was working in, uh nfts and the dota persona has just been the my online persona ever since okay you know naval ever always says uh the best uh the best outcome is to be famous but nobody knows you i i understand and listen you got good hair so yeah you know i'm not i'm not judging all right let's get into it all right yeah so let me talk about paperclip so paperclip here I can show you this dashboard.

2:37This would be the dashboard that I use to manage Paperclip itself, which is pretty involved. So let's take a step back. So the idea with Paperclip is it is a tool for the tagline is zero human companies. Now, I think that is maybe a bit aspirational. I'm sure you guys have seen tools like Pulsia, where you give it your credit card, it creates business ideas for you. And Paperclip is designed to have a little bit more control right now because it's designed for work that you're really accountable for doing. The idea behind Paperclip is that you're really managing your business goals. So it's not, you know, on one end, you might have something like Pulsia.

3:21It's totally automatic. On the other hand, you might have like an AI coding tool where there's a bunch of tabs open and you're managing pull requests. With Paperclip, your idea is you're going to manage business goals. You define your goals, you hire a team, and then you approve what they're doing and they go and they work on it. So an interesting thing about Paperclip is that it's bring your own bot, right? So you're already starting to see the labs come up with project management tools. You have OpenAI, you have Anthropic. With OpenAI, they have Symfony. With Anthropic, of course, they're releasing new project management tools every day.

3:54with with paperclip um you're able to bring your own bot if you use cursor cloud if you use open claw if you use um clod cli like i would say that all the programmers that i know or all the like entrepreneurs that i know are using lots of different models because they all have such a different personality and that's one of the key things um about paperclip is that you can bring your own agent uh into the system so yeah greg one of the things i thought we could do is we could go to your idea browser. Yep. And we could look here and we could just try to decide something from the idea browser that we could try to use as a case study.

4:37So I don't know, what's one of these that you think might be a good option? If you, I mean, there's a few really interesting ones. Like I like the, you tell me, I'll give you two and you pick. So scroll up. Yep. So if you go down, there is a finance app that builds money habits in three minutes a day. Okay. That's one option. And then the other option, if you scroll down. Yep. The flight compensation recovery agent for frequent flyers. Both of those are great ideas. Okay, nice. Yeah, let's try a finance app that builds money habits. So one of the things I noticed actually when I was testing this out, if you just right click, you can copy like pretty much a good description from this more than more than this.

5:30So here we can come in here. Like, what do we want to name our app? We'll name it. Moolah. Moolah. Okay, great. And what is this company trying to achieve? Right. So this text box, I suppose, is a little small, but we'll just paste it and we'll come back and we'll look at this in a second. After you write down what you want your company to do, you create your first agent. So Paperclip right now best works when it's on your local machine and especially if you already have something like CloudCode or Codex installed. I think these tools are really important for really any agent-driven entrepreneur.

6:06I'm guessing pretty much everyone in your audience has at least like CloudCode or Codex installed, right? We recommend that you use one of these two for your CEO, which would be your first agent, but we do support other agent types, right? You can use open code, which means you could use really any model that you have on open router. So you've probably seen open router. They have like, what is it? Leaderboard that kind of shows the most popular, yeah, the rankings. They have the AI model rankings. So you can use any of these models with Paperclip. I think you're going to get the best results with Paperclip if you use a frontier model for your CEO, but then maybe for other tasks, you can get away with a cheaper model.

6:47A hot tip is OpenRouter actually often has free models. So there's been this Hunter Alpha model was free for a couple of weeks, the StepFlash free. So if you have work that can be accomplished by a free model, right, which would be sort of, you know, not all work can, you can use it and get free tokens. Also the Pi coding engine, whatever. So it basically depends on some of these existing agents to operate. This one's actually going to run on my local machine. So we'll just go next. And do you recommend doing local or how should people think about this? Yeah, so right now, Paperclip mostly works when you run it locally.

7:33We're working on doing a cloud deployment. So either you can self-host it yourself because it is open source, or eventually we'll provide a hosted solution for people that want to do that. but i think local is really how most people are doing a lot everything with paperclip right now yeah cool so the first thing we want to do is uh give our agent something to do hire your first engineer and create a hiring plan so you are the ceo you set the direction for the company hire a founding engineer write a hiring plan break the roadmap into concrete tasks and start delegating work so you can edit this if you want um you know maybe your business doesn't need an engineer.

8:10Maybe you're like doing marketing sales. You can sort of adjust this depending on what you want. But, you know, having a founding engineer is helpful because almost every sort of online business needs to do a bit of coding. So we'll click next here. Okay, we've got our company, our plan, and it kicks off. So what you have here is along this sidebar, we've got different companies. Paperclip is made to run as many companies as you want to. So I'm running Paperclip itself in here. I'm running Forgotten Ruins from here. I've got a couple other companies that I'm like, you know, different apps that I'm working on.

8:49And here we've got our new entry for the Moolah company. You can see right here, we've got this dashboard. Part of the thing with Paperclip is that it will track the monthly spend and it tracks all of the work your agents do. I mean, I built Paperclip because I was using CloudCode to build my companies and I would have 20 or 30 CloudCode windows open all at once. And I couldn't remember what any of them were working on. I would set them to run over the weekend and I'd come back and I had no idea what anyone did. I couldn't remember what they were supposed to do and they had spent all my money on all their tokens.

9:28My budget was busted and I didn't know what was actually accomplished. And so Paperclip is designed to solve those problems. It actually tracks your monthly spend that you have in tokens. Everything that happens is done through issues. Right now, we have this first project here and it hires our first engineer. These things sort of take a while. Here we go. We can see we've already got our hiring plan installed. Immediate hire, founding engineer, you know a roadmap breakdown etc etc you know what it actually is for this call doesn't really matter but you could look in here you could edit it give it feedback give it comments so um what happens in paperclip is your agents um have tasks that get created and then um they work on them i mean that's it's as simple as that right and uh paperclip sort of orchestrates this idea that each agent, each task can only be handled by one agent at a time.

10:34So you're not going to get this kind of conflict of agents stepping on top of each other. Okay, look here, we've got something in our inbox. Okay, hire agent, our first engineer. There's a request from our CEO to hire an engineer. Now, as you get more comfortable with paperclip and how your agents are acting, you can turn this off and let your agents hire agents automatically. But for sort of these like defaults, we've said, you know, a lot of people are getting started their first time. We want to approve this manually. So we just approved that our CEO can hire his first engineer. So is that sort of the big mindset shift with Paperclip?

11:15Instead of, you know, thinking of jobs to be done? You think of hiring agents? Yeah, I think one of the things, instead of thinking about like the details, the goal of paperclip is that you're thinking at a higher level. Like you think of yourself as like the board and you should be talking to the CEO and saying, here are high level goals. Here's the MRR that we want to reach. Here's like what we want to achieve. Now I'm asking you to make a plan and institute what needs to happen within the company. So you can see here, we've hired this engineer and now we've got like five live tasks actually happening here.

12:01So let's look in a bit more deeply and see what's going on. So if we click on the issues, you can see we've got more issues. Set up the scaffolding, continuous integration, do progress tracking the user auth and onboarding build a core daily loop right so the agent right now is uh working on these things the engineer is um now i can see right now that the engineer is actually just working on one of these um so i might go in here and i can up the concurrency so whatever there's advanced settings in here where you could say you know run four engineers at once and it's a bit in the weeds but uh so wait going back to that because i you know I know it's the weeds, but the configuration, like should we spend any time in here?

12:47Sure. How should we be thinking about this? Yeah. So one of the key things that you want to be able to do with your agents is you need to configure them with a bit of persona. So the paperclip right now gives you a default persona for your CEO. So if you look here, it says, you know, you are the CEO. Here's how we expect you to use memory. We use a memory system by default that just stores your memory in files that was designed. It's using like the para memory system, which is like, I think, Diego Forte's invention that this skills from Nat Eliason. He runs the Felix bot. So it gives your CEO a default memory.

13:35Now, agent memory is a really involved topic. There are a lot of different choices, which I'm happy to talk about. But this gives you a simple default. And then you give us some safety considerations, right? Don't exfiltrate my secrets. Here are some other things that I want you to look at. One of the key things that you want to look at is the heartbeat. And here's the CEO's heartbeat checklist. So when you think about AI agents, Here's a tip about using AI agents with paperclip or anything. Have you ever seen the movie Memento? Of course. Yeah, yeah. So for people who haven't, Memento, the Memento man, he wakes up every morning and he is like really capable, but he can't remember where he's at or where he's doing.

14:22And he has to leave himself a little notes. Right? Your AI agents are Memento man. They wake up. They know how to fight. They know how to drive. They know how to like just take care of themselves and spend money, but they don't know who they are. They don't know where they are. They don't know what they're supposed to be doing. And so what you need to do is actually write down, you know, give them little Polaroids or write tattoos on their arm on who you are and what you're supposed to be doing. And so the heartbeat basically means like when you wake up as an agent, what are you supposed to do?

14:56And this gives you these instructions. So fetch Agents.me, confirm who you are, learn who you are, read today's plan, look for your assignments, check out the work, and then break it down into tasks and extract your memory and then you're done, right? So these are your responsibilities. I would say that, you know, Paperclip has been out about three weeks and we are already improving on this on a daily basis. You know, one of the key things that you always have to do with your agents is you really want to set up a way to kind of evaluate the quality of the prompts that you're giving them. And this is really like a new field that a lot of people haven't, it hasn't been figured out yet.

15:43We'll be building more tools in Paperclip for helping you do that. but really how a lot of people just do it practically is when it does something you don't like you come in here and you say rule like like make sure you remember to set you know like if it's forgetting to set like a win condition like make sure you set like a success condition for every task and ask the developer to have QA review it, right? Like I might be using this and I realize he never defines what success looks like and he never asks QA to review it. And I'll literally just like add that line in here and then save it. And then, you know, typically you'll get a bit better results, which by the way, I think is something that we should have our CEO do.

16:32So I'm going to ask him to hire a QA person to help review the work. So that's another thing. So this is another way that you use paperclip is whenever you have an idea for something, just ask. The CEO knows how to use paperclip very, very well. And if you're having trouble, if you want to set up a new project, you want to set up, you improve your organization, you want to make a plan, you don't, you need advice, you just ask your CEO to do it. So I'll create this issue, let him work on that. Another key way that you might want to configure your agent would be with skills. So, you know, if you've tried to use, you know, OpenClaw, for example.

17:22Okay, so let me talk about OpenClaw for a minute. OpenClaw is insanely powerful, right? It's such a beautiful piece of software just taking the internet by storm. It's really the first time that, you know, if ChatGPT was the first time people realized that AI could talk, OpenClaw was the first time people realized AI could act, like that you could actually have AI act on your behalf. I would say that like Paperclip is aiming to be the third moment where you realize AI can do real work that you're accountable for, but that's the future. Like with OpenClaw, one of the things that's amazing about it is that it can just do so much for you automatically.

18:07But the problem that people run into is like, if you use OpenClaw for two or three or four weeks, you start to realize that it kind of like falls apart. You don't really know what it's doing. You don't really know what it can do. It breaks and you don't know how to fix it. and so paperclip kind of paperclip can use open claw as an agent but paperclip is designed to give you a bit more accountability into what everyone's doing you can break your agents up into different um personas and you can also give them different skills um the best way to find skills would be uh skills.sh maybe not the best way but it's one of the most popular have you used this before yeah i my my question here is how do you know you're not installing skills that are you know have bad stuff in it it's a problem yeah it's it's a it's a real problem and it's something you have to be careful with i don't think anyone solved that um one of the things that skills.sh has is they do have some degree of security audits.

19:16So they have these badges. So I think that that's as best as you can do probably for now. And I mean, you can, just because it has a lot of GitHub stars doesn't mean it's, you know, 100 % secure, but it does give you directional kind of it's a data point right if if it has if the remotion you know skill has 100 plus thousand stars for example chances are it's likely that there's value there that's right that's right so what we can do if we wanted to if we did want to make a video about this right we can copy this here and we can go to paperclip there's a couple things that you could do you can um your skills are organized by a company.

20:06So we should be able to paste this here. Or if you can't remember that, you can just ask your CEO, like add the re-motion skill or whatever. But we don't need to do that right here. We can just click add. Yeah, there we go. So now we've got the re-motion skill installed into our company. And this actually keeps track of, like you can check for updates if the skill gets updated, et cetera, et cetera. You know, something that I might often do, like I might do is I might say, you know, hire a video editor and give them the remotion skill. And I would paste the URL and the CEO knows if you sign the CEO, he knows how to do this.

20:52For the sake of demo, we'll just wait because we don't need to do that this second, but I'm just telling you that's that's that that's how you use paperclip is just ask the ceo to do things for you so i have a question i have a question about that specifically so you know the re-motion skill you know i've used it and i've gotten sometimes i've gotten good results and sometimes i've gotten bad results and the way i've gotten good results the best possible results is giving it the most possible context the most possible references um and so my question is you know if i'm going to hire a video editor and give them the remotion skill how do i ensure that this is the you know a top one percent video editor using paperclip so you you can never get around the idea that you have to provide that context ultimately like paperclip is not going to um on its own write that context for you.

21:52That said, it does have tools that can help. So when you have issues, you can have them. We can hire our QA agent. In your issues, we do have documents and we're going to add features to do a better job at like organizing the documents, organizing the artifacts that you have within Paperclip. We'll do that. But of course your agents are able to access pages anywhere you want. So for example, on the paperclip website, I have a brand guide. Yeah. The brand identity. So when we're asking this question, which is like, how do I make sure that my agent makes a video that's good? One piece of that is like, do you have a brand guide yet?

22:42And you can ask your agent to do this, right? So I would say for kind of moolah, we haven't necessarily gotten there yet, right? Our agents are still kind of working, but eventually what you would do is you would create an issue, create me a brand guide, and it would give you the URL and you would look at it. Then the other part would be, as I might say, well, why don't we do this, right? Let's say, let's plan a 60 second remotion video about moolah showing the key pitch that we could share on twitter and we'll sign this to the ceo for now because we don't have a cmo but maybe we should maybe it should be in another project but whatever we'll we'll get there.

23:32So, um, I would say make a plan for this. Don't, uh, make it yet, but we need, uh, a video demoing the product and showing the pitch. It needs to be animated. Needs to be animated and pretty. Agents don't really like care about typos. So you don't have to care that much about it. So whatever, we'll let that run and we'll come back a bit later. One of the things you'll notice is that our monthly spend right now is$0, even though we've been doing all this work. Part of that is because I'm using my Codex subscription or Claude code subscription. Actually, I'm not even sure which model we're using for these.

24:19It might be Claude. Oh yeah, look at this. Our agent, yeah. So our CEO is using the Claude Opus and it's, you are able to use your subscription as well as API credits if you want. And that's why you're seeing here that the costs are zero. It's a little bit of, you know, you're still using your kind of inference for your subscription. But if you hooked it into something like open code or were using like pure pay per token values, you'd get kind of real dollars here. Okay, we've hired our video editor. And let's check and see if he was given the skills. Yeah, so there you go. He has the re-motion best practices skills enabled our ceo is still working on uh planning the video so we'll just let him kind of kind of work doesn't seem like you have the context of firing a ceo firing an engineer firing a qa agent is that correct no so yeah you you can yeah you can come in here to configuration and uh let's see oh where's oh let's see here i mean i hate to be that guy but no right here we got terminate you can terminate yeah okay well don't don't do it but you know he has had a chance yeah yeah give him a shot give him give him a 30-day cure period at least right yeah so um some of the other things that you uh can do with paperclip is this idea of shareable companies.

25:58So let me take a step back, which would be, it is not that clear how to structure your company in a way that operates really cleanly and creates good results, right? Everybody who's sort of tried to one-shot a new startup with AI, you realize like it's super fun for the first half an hour and then it just kind of falls apart. There are a lot of new patterns that are coming out around like agentic design patterns. OpenAI has a blog post where they call it harness engineering, but it's just a fancy way of saying like ways to have your agents interact with one another that will help you get better results than just trying to quote unquote one shot results.

26:45So these patterns are still being developed. I would admit that I'm even learning them as I go. A really basic one would just be like having a process within your organization where after the engineer creates something, he asks QA to QA it. This is especially important for like web apps where you need to like click on something and make sure that it actually works. So for example, you know, in here you can like at mention the CEO or at mention a project, you know, like the onboarding project. And this was like a really, you know, you see it all the time, but for whatever reason, my agents had a hard time coding this.

27:26Funny, I can even see here that the, that the pills have a different height and that really annoys me. Um, so like what I might do is I would go into paperclip and I would actually post this screenshot here. And I would mention pill. Fix it, right? And I'm going to assign this to the clod coder for the paperclip app. And I'll say, you know, when you're done, pass it to QA to get eyes on it and verify it works here with a screenshot. Now, listen, if you're going to run a zero human company, you cannot be managing your apps at that level, right? Like I'm actually acting not even as a board member or CEO.

28:27I'm actually acting as like a design PM or something. And so there is kind of this like structure that needs to be built where you should have a design agent who's already looking at the website and detecting that stuff. And so that way you don't have to be kind of micromanaging your agents. And I would say we'll get to that, right? You basically have to make sure that you're ready to spend the tokens to do that. And you also need the evals to make sure that your designer sort of shares your values in terms of how you want the app to work. I would say that like even the models today, the best like GPT 5.4 or like Opus 4.6, some of their taste is still not quite there.

29:16And that's why, that's where a lot of the secret sauce is. That's where you're actually going to write your own skills. And that's where, because you're basically going to impart your own taste. Like AI can do everything except know your values. And so you actually have to become more aware of your values and find out how to communicate them back. Which is, even in a pre-AI era, the concept of a good leader, of a good manager, of a good CEO, of a good founder, is very much someone who can clearly communicate their values and taste, right? That's right. That's right. So not much has changed except the vehicle to doing it has changed.

30:05Instead of hiring employees, you're hiring agents. I had a question before we move on. I had a question if you go back to the Paperclip project, which by the way, it's so cool that you use Paperclip to build Paperclip. I expect nothing less from you, obviously. But you were talking about it's important to hire an agent for this and an agent for that. How many agents do you have for the Paperclip project? Not that many, actually. So this would be the org. It's actually a little hard to see here. And yeah, I would say that it is growing all the time. So we've got the CEO reporting to him would be a CMO, a UX designer and a CTO.

30:57And I use different agents for different reasons. I would say that the cursor coder and the clod coder would be my kind of workhorse. They're using the frontier models on my subscriptions. Then I have a QA engineer who has both the Claude browser installed as well as there is a skill called agent browser, which is it basically gives your agent access to a web browser in a way that is a little bit faster than using Chrome. Like if you've used Chrome with Claude code before, I'm sure you've seen it like pops up a Chrome window and like takes over your computer. And if you click on it, you mess it up.

31:38And so that's the QA. I also have, I'm starting an evals engineer. So evals would be basically just a way saying like, how do we start to kind of like, how do we do performance reviews on our agents and be able to go back and reflect on what they did? Paperclip is going to be building tools where you can look at your past issues, look at the feedback that you've given your agents, and your agents will be able to learn from that, right? If you find that you're giving the same feedback over and over, we're going to learn that, and you won't have to make the same mistake twice. We've got a UX designer who kind of has context around what I like for design.