# Open Claw Runs My $11M Business: How To Get Rich In The New Era Of AI Agents (Even As A Beginner!)

https://www.youtube.com/watch?v=MDtMwKcx_4E

[00:00] For less than $100 a month, you could have a team of agents working for you at all times.
[00:06] I actually have my AI agent 24/7 negotiating all of my brand deals.
[00:09] Chris Camillo, he's like a investor.
[00:11] He said, "You can probably make half a million dollars a year doing this."
[00:14] Yes.
[00:16] I think there's a lot of opportunity in learning how to create a good employee and then selling those employees to companies.
[00:22] I think that'll be a massive trend.
[00:23] What is the thing that you would advise or suggest that someone listening at home goes and does?
[00:30] In this video, we'll create an agent, create one together.
[00:33] Like we can come up with an idea and I'll show you exactly how to set up a cron job, which is just an automation, and then we can test it.
[00:39] 12 months ago, we had a conversation that led to hundreds of thousands of people vibe coding and building their first app with AI.
[00:50] And so, where I want to begin today for the person that's sitting at home, if they watch this video to the end, what are they going to be able to do as
[01:00] a result that they can't currently do now?
[01:04] Yeah, I think my main two goals for this episode is one, um to explain what an AI agent is and kind of conceptually.
[01:16] I think it's really important for people to actually understand what's going on, at least a little bit beha- um behind the hood in terms of what an AI agent is.
[01:23] And then we're actually going to create uh an AI agent, right?
[01:27] We are going to build an AI agent that will be useful to me.
[01:30] Uh and uh I'm going to go like from setting up an AI agent to actually have it running so that it does things uh automatically and every day.
[01:42] That's like really useful for me and my business.
[01:43] And I think it is the most important thing for people to understand right now in the world of AI.
[01:48] Mhm.
[01:49] You know what? You said you said it's the most important thing for people to understand right now in the world of AI.
[01:54] If we rewind 10 months ago when we had the first conversation, everything that you were speaking about
[02:00] and where your energy was was with building apps, this new ability to build software like the non-technical people using AI.
[02:10] Can you just get people up to speed?
[02:13] Like what has changed from 10 months ago to where we are now where everyone's talking about AI agents?
[02:23] Yeah.
[02:23] So, um the term vibe coding uh was basically just using agents to build apps, right?
[02:32] That's all it was, right?
[02:32] You have really smart AI models, you type in your idea for an app, and the AI agent generates all of the code for all of the files.
[02:39] It can edit files, it can delete code, it can do anything uh within like a directory, and then it can build an app, and you can immediately start using the app, and many people have actually made money selling software that they've completely vibe coded, right?
[02:57] Without any coding experience.
[02:58] Now,
[03:01] it turns out that those same coding agents like Claude code are incredibly good at just doing general tasks, right?
[03:09] If you think about any reason or all the software that you use, uh you know, whether it's like a CRM, whether it's Notion, all these different things, it's it's usually to accomplish a goal.
[03:16] So, people were vibe coding apps to accomplish a goal.
[03:21] People started realizing that you could actually just give give these agents skills, and the agents could actually do those tasks directly.
[03:27] It didn't need any software in the middle to do it.
[03:30] The agents could just start doing all of the tasks, whether that's just handling all of your email, right?
[03:36] You know, uh as a content creator, you know, you probably get reached out to by many people per like every day, probably many companies reach out to you, and I actually have my AI agent 24/7 negotiating all of my brand deals, you know, and and I do actually don't even need to be in there.
[03:55] And I used to have a really good manager who would manage my sponsorship deals.
[04:00] And uh I basically gave it all of the
[04:02] data from the previous brand deals.
[04:05] And now the agent just directly negotiates with companies on my behalf.
[04:10] And so that's where this is moving across all industries.
[04:13] And that's why you're hearing about um, you know, a lot of companies moving uh spend to creating AI agents.
[04:22] Yeah.
[04:24] You know, you said that's where it's moving across all industries.
[04:26] I remember when we spoke previously you gave the example of like uh like Peter Levels and these vi- these apps that he was vibe coding.
[04:35] And I think at the time he was making like $300,000 a month vibe coding apps.
[04:40] If we were to move the conversation forward to AI agents and how people are using things like Claw, Open Claw, or Manas, or even Perplexity Computer.
[04:50] Can you just share some examples of how people are using almost like these AI agent workflows in their businesses, or to make money,
[05:03] or like in their professional lives.
[05:05] Can you just share what you've seen?
[05:08] Yeah.
[05:10] So, I think obviously the biggest uh use case now is still coding, right?
[05:15] Because we have There's been a lot of like people are just building internal tools all day long that are super useful for their business.
[05:23] And the reason why it's so exciting right now to get into these general AI use cases is because it's brand new.
[05:29] People are starting to realize it can do like general things beyond building apps like 2 months ago or 3 months ago.
[05:33] So, it's brand new.
[05:36] And so, all of these use cases are are brand new.
[05:39] Some of them aren't perfect, but we're reaching a point where they're becoming perfect.
[05:42] Now, uh the one that's probably Open Claw was incredibly unique in that it was an agent that just ran 24/7 on your computer.
[05:54] A lot of people are using it for project management, and a lot of people use it for daily recap.
[05:58] So, it has access to my email, it has access to linear, it has access to Notion,
[06:05] And my calendar.
[06:08] Basically, all of the tools that I use to run my business, Openclaw has access.
[06:13] And so, if Openclaw thinks that there's something important that it needs to tell me, it'll just do it.
[06:19] And that is kind of what we'll talk about later today is there's levels of autonomy, right?
[06:23] You can create on one side of the spectrum, you can create an automation, which is like a trigger.
[06:26] So, if something happens, then the agent will do that.
[06:33] But now, what Openclaw did is they released a feature called heartbeat.
[06:36] And the heartbeat is basically says to Openclaw, it says, "Wake up every 30 minutes and decide if you want to do something."
[06:43] Right?
[06:43] The agents are starting to decide when it needs to do stuff, which is a whole new paradigm.
[06:47] Right?
[06:47] And and so, the level of autonomy is increasing over time, and the and the length of time an agent can work productively is increasing over time.
[07:00] So, they're getting more autonomous, and they're able to work on harder and longer tasks.
[07:05] And I think by the end of this year,
[07:08] we're going to be measuring how long they can work in days, not in hours, which we're currently up to like 6 or 7 hours that agents can work.
[07:15] Yeah.
[07:16] So, it's a really exciting time.
[07:18] Okay.
[07:18] So, okay.
[07:18] So, you mentioned that it's a really exciting time.
[07:20] You mentioned that we're we're kind of entering this new paradigm.
[07:25] I remember a few months ago when Openclaw kind of shot to prominence, and you're seeing everyone, at least like on Twitter, like in these tech circles, they're buying like Mac minis, they're you know, they're getting set up in a weekend.
[07:41] Everyone's like going deep with it.
[07:45] And from what I've heard from you, and I think you have a really interesting vantage point because you're deep in this world, but your background wasn't in being a coder or like a software engineer or like a tech a deep tech like insider.
[08:02] And so, my question is for you, if we project out, especially as this
[08:08] technology becomes more and more autonomous, you talk about heartbeat.
[08:12] What do you think it's going to be able to do if we project forward a year from now, 2 years from now?
[08:20] Like, what does this mean for people?
[08:22] What does it mean for people?
[08:23] It means that basically everyone within a year will be able to have a team of AI agents working for them or alongside them or with them for, you know, I think it'll start off a little bit more expensive.
[08:38] It might be $200, $300 a month, but as the models get a little bit better and for cheaper, I think for less than $100 a month, you could have pretty useful AI agents a team of agents working for you at all times.
[08:55] And I think people who build businesses that have like true value and you are able if you're able to actually leverage AI agents to like help propel your business,
[09:08] those people are just going to be really successful.
[09:09] And so, learning how to think in systems will be probably the most important thing people can do, and you'll AI agents will feel more and more like talking to a person, right?
[09:24] When you hire a new employee, you have to like train them, you have to teach them like what your company does, you want to show them what uh a good outcome looks like.
[09:31] Maybe you hire a researcher for your podcast.
[09:33] You you kind of want to train them a little bit, and then you want to like set them loose, and eventually they'll be able to just do tasks for you like an employee.
[09:43] And so, I think over the next two or three years, it'll just feel like hiring an employee.
[09:47] Anything that you can do on the computer, you'll be able to just hire an AI agent to do the same thing.
[09:52] And so, instead of paying them per hour, you'll be paying for the tokens, for the the AI tokens, basically.
[09:58] Yeah.
[09:58] You know, it's interesting cuz I was listening to this um podcast with Chris Camillo, he's like a investor.
[10:07] And one of the things he said, he was
[10:08] talking about AI agents and Open Claw specifically, he said, and he was talking about AI automation, like AI agents automating workflows and tasks.
[10:19] He said, you can probably make half a million dollars a year doing this.
[10:22] There are people doing it right now.
[10:25] He goes on to then say, "Historically in the dot-com boom, the gold rush was trying to pick the right company to work for.
[10:33] You pick the right company and you have a chance of being rich."
[10:34] He said, "This is a very different gold rush for this generation.
[10:39] It's not about who you choose to work for, you have to do it on your own, but the opportunity is near infinite with AI.
[10:46] The difference is the window is very small.
[10:49] It's like right now, it's this year."
[10:53] I'm curious cuz it sounds so hyperbolic.
[10:56] I'm curious to what extent do you agree, disagree?
[11:02] Yeah, I I I will say for the first part, you know, I'm in San Francisco right now, and uh everyone who worked for
[11:09] OpenAI or Anthropic, you know, when they IPO, every single person is going to make like three to 20 million dollars.
[11:16] Uh so, like to work for one of the big AI labs is a massive opportunity right now, and I think you can actually make a ton of money working for the right AI startups right now.
[11:25] I think the growth is insane.
[11:26] So, I will say that as a caveat, but I do agree with the sentiment around how there's a ton of opportunity to work on your own.
[11:32] Yeah.
[11:35] That's It's It's so interesting.
[11:36] You know what? We're going to I want to get into AI agents specifically and how people can go about kind of building that first one.
[11:44] Before we do, in that quote, Chris Camillo mentions right at the end that there's like this window of time right now.
[11:52] And the way that he positions it and frames it as if is is as if it's like a limited window of time.
[11:59] Do you agree with that sentiment that like in the next year or 18 months or 2 years, there's really like a unique opportunity to capitalize and use this technology?
[12:11] either to build businesses or even for someone who's working at their job that wants to like automate workflows?
[12:18] Yes, I do. And I don't I really try and to give people urgency in other ways other than like you have 2 years to capitalize on this or you're screwed.
[12:28] You know, there's a meme going around in San Francisco called escaping the permanent underclass.
[12:32] I don't like that people talk about this.
[12:37] There's There's this idea that AI's going to get so good you have 2 years left to like acquire money before it's just a bunch of elites who control the AI and then the rest of the the permanent underclass.
[12:50] And it's it's a meme and I I think you should avoid that type of thinking.
[12:52] You should find other ways to motivate yourself like AI is here.
[12:56] Creating AI employees is is kind of difficult right now.
[13:00] And when things are difficult, that's where all the opportunity is.
[13:03] As soon as it is super easy, as soon as you can just go to a site and click yep, hire a new employee and then boom, you have this which is
[13:11] coming, you know, and that might be 2 years away.
[13:15] As soon as that happens, right, all the value kind of goes to the companies who create those AI agents or the model providers.
[13:21] And so right now there's this moment of time where there's friction and it's hard and that's what companies are willing to pay for.
[13:29] So, in terms of creating AI agents or automations or starting a business, I do think it's a really good time right now to get in and you can make a ton of money doing it.
[13:37] Are you screwed if you don't do it in the next 2 years?
[13:41] I'm not sure. Um I I I wouldn't think like that.
[13:43] Yeah. I like that. It feels like a a balanced uh perspective.
[13:47] You know what, uh Riley, cuz just context for the audience, you and I had a conversation um about a week ago, and during that conversation, you said something that the moment you said it, I wrote it down so that I could bring it up today.
[14:01] And basically, what we were talking about is most people, when it comes to AI, they're familiar with like the LLMs, right?
[14:08] They're familiar with ChatGPT.
[14:10] They might, you know, started using
[14:13] Claude um recently.
[14:14] And you said something, you go the trillion-dollar battle that's going on between OpenAI and Anthropic.
[14:23] Anthropic are the creators of Claude.
[14:25] And it's interesting because it's It feels like every week there's a new product update, which is like a a sizable leap forward.
[14:32] I I I I want to actually get into some of these AI agents, but the the amount that this is changing like week over week, and I think about what we have with like Manas, we have Perplexity Computer, you have uh Claude Code, you obviously have Open Claw.
[14:52] Like, there's so many players, and it's developing so quickly.
[14:57] Can you give people that context of where we're at right now, and then we can we can start showing people how to build?
[15:09] Yeah, so Okay, so you named a bunch of tools, right?
[15:11] There's a lot of different tools right now on the market, like Open Claw.
[15:15] You have Claude Code and Claude Co-work.
[15:18] You have Perplexity Computer, you have Manas, and what?
[15:24] And I'll I'll I'm going to explain this in a little bit when I I I when I go through I have kind of a presentation prepared as kind of a precursor to building the the agents.
[15:30] But what separates all of these tools from Chat GPT is it's like an AI model that has access to a computer.
[15:36] And um And what we're seeing now is AI is getting really good at doing every single thing a human would do on a computer better and faster.
[15:48] Which is has massive implications.
[15:51] If you think of all the knowledge work, right?
[15:53] It is on a computer.
[15:54] If you work at Microsoft, you're doing most of your work on a computer.
[15:57] You're handling email.
[15:59] You're using maybe a suite of like seven or eight different tools to like accomplish tasks.
[16:04] AI is uh you know, Anthropic just acquired a company called Percept, and Percept is a computer use tool.
[16:11] And if you you can actually test it.
[16:13] They have like a Vi agent that like if you give it
[16:15] a task, it'll just like quickly control it literally controls your entire computer and completes tasks.
[16:21] And so what we're seeing is AI agents getting smarter than way smarter than the average human, and they're also getting better at controlling a computer than a human is.
[16:30] Um and so in terms of like where we are now, all of these different tools that I mentioned use a computer to accomplish tasks, but they all do it in a slightly different way, which I can break down in the next section if if you want.
[16:45] It is kind of a longer conversation.
[16:46] Yeah.
[16:48] And so you know what? Cuz I I uh We let's go ahead and and and show the presentation.
[16:51] And even while you're you're pulling it up, I'm curious cuz you run a company and a business, how many of these different agents that are Like how many do you even have running as we speak and like doing tasks for you?
[17:10] Okay, so you just heard Riley talk about the power of AI agents.
[17:11] And listening back to it, it reminded me of
[17:16] this new tool that I've been trying lately called Gamma Imagine.
[17:20] And so to give you the backstory, I had this cool idea for a visual that I wanted to put at the end of our Instagram Reels.
[17:28] And so the problem is is that I'm not a design person.
[17:31] I'm not a trained designer, but often I have to come up with visuals.
[17:36] And I hate that like manual finicky process where you're working around with these different templates and tools and it never works and it takes so much time.
[17:46] And so that was the process that I was stuck in and it was incredibly frustrating.
[17:49] But luckily, in this world of AI, I thought to myself, there must be a better way.
[17:55] And so to take you guys behind the scenes for a second, I took the idea that I had for the Instagram Reel.
[18:02] I simply typed it into Gamma Imagine and within seconds it produced stunning visuals like this.
[18:08] And another thing I love about using it is that it can come up with iterations and improvements really fast.
[18:15] You just type in the improvements, the feedback
[18:17] that you have for it and it produces it on the spot.
[18:19] And so if you're in the position that I was in, where you're completely fed up with the manual design process and how time-consuming and frustrating it can be, I highly encourage you to go to the link in the description and try using Gamma Imagine.
[18:36] Let me know about your experience.
[18:38] I know you're going to love it.
[18:40] Okay, so I'm going to share a story from my early days of building my company.
[18:42] So back when I started, I wanted a website for the show.
[18:44] But when I tell you building one was such a struggle.
[18:49] I would watch design tutorials.
[18:51] I tried copying other people's websites that I was inspired by.
[18:56] None of it worked and the worst part is it took hours, sometimes weeks for me to do this and it didn't look any good.
[19:04] And that's why I'm excited to share with you the sponsor of today's episode, our friends at Hostinger.
[19:08] And Hostinger makes it so easy to build a beautiful looking website.
[19:11] All you have to do is
[19:17] you open their website builder, you describe how you want your website to look, and within minutes you will have a beautiful looking website.
[19:24] And after you launch your site, let's just say you wanted to start sending emails, you can also use Hostinger for that.
[19:30] You don't have to subscribe to a new email software.
[19:34] So Hostinger has emails, automations, website, and AI tools all built into one.
[19:41] And so you're probably thinking, "Okay, that sounds good, but there's no way that I can afford all of this."
[19:47] Well, that's actually the best part about using Hostinger.
[19:49] The price is designed for people who are just starting out.
[19:53] And so if you have a business idea, or let's just say that you're looking to build a beautiful personal website, go to the link in the description and start using Hostinger today.
[20:03] Thank you to Hostinger for partnering with us on this.
[20:06] Let's get back to it.
[20:07] We're a software company and every part of our workflow is run by AI agents.
[20:10] So from building from generating the code to checking the code to
[20:20] basically automating like customer
[20:22] support, it's 95% automated. Like
[20:25] whatever we can
[20:28] replace with an AI agent, we do. And
[20:30] it's really cool to see. And now as a
[20:32] marketer, a lot a more and more, like
[20:34] right now all of my DMs and outreach to
[20:38] people to test kind of a new software
[20:40] platform we're releasing is done with AI
[20:42] agents. And so at any given time, like I
[20:45] think I have like
[20:47] like right now I have like seven or
[20:48] eight different AI agents working every
[20:51] single day just for me personally, but
[20:54] our team like we spend
[20:56] we spend six figures a month on AI
[20:58] tokens just for the AI agents that we
[21:00] use internally. So
[21:01] that gives you a sense of scope.
[21:03] >> And and you would say it sounds from
[21:05] your
[21:06] experience
[21:07] the most useful use cases right now is
[21:11] customer support and marketing.
[21:15] >> Yeah.
[21:16] Um yeah, and then and then uh just like
[21:17] executive assistant tasks as well.
[21:20] Um
[21:21] and just just organization, you know,
[21:22] like I get hundreds of emails a day. I
[21:26] get so many YouTube comments every
[21:28] single day across my videos. You know, I
[21:30] have an agent that just analyzes all of
[21:32] the comments for I have it as a weekly
[21:35] workflow. And we'll talk about creating
[21:37] an automation or a cron job that's that
[21:40] it's basically an agent that does a
[21:42] super long task. And so yeah, that
[21:44] But the point is here is like these
[21:46] agentic workflows aren't just like plug
[21:48] and play. They're they're very catered
[21:49] towards They're very they mold into your
[21:52] existing workflows. They use my tools,
[21:55] you know, and you may use a different
[21:56] tools. Like I specific I use Notion,
[21:59] Canva, and Linear like every single day,
[22:02] you know, but you might use a completely
[22:03] different stack. But the point is these
[22:05] agents can mold to whatever tools that
[22:06] you use. Um
[22:08] and that's what makes them so useful.
[22:10] >> Yeah. Just before we get into it, cuz I
[22:12] know you mentioned that your company is
[22:14] currently spending six figures a month
[22:17] in like just the tokens to run these
[22:20] agents. And I and I want to make this
[22:22] clear for people. What is just as an
[22:26] individual looking to get started on
[22:30] this, what would be my setup cost? I
[22:34] know people are talking about like Mac
[22:35] Minis. You have to buy the $800 Mac Mini
[22:38] in order to Like how much would you say
[22:40] for a beginner looking to just start
[22:42] building their first AI agent to handle
[22:44] one workflow, what would the cost of
[22:46] that be?
[22:48] >> To handle one workflow, I mean, you
[22:50] could um
[22:51] you could use Claude Co-work or you
[22:53] could use a hosted version of Open
[22:55] Claude for 20 bucks a month to handle
[22:57] one specific workflow. If you had one
[22:59] workflow that ran once a day, you know,
[23:01] that that may cost you like
[23:03] 20 bucks a month. Uh you may cost you 30
[23:05] bucks a month, depending on how long the
[23:07] task is. Maybe you have a really
[23:08] long-running task
[23:10] that spins up sub agents, which is
[23:12] actually a lot easier than it sounds.
[23:14] Um you know, that may cost $50 a month.
[23:17] But it is not that expensive and you do
[23:18] not need to buy a Mac Mini.
[23:20] You can buy a Mac Mini and I encourage
[23:22] people, if you're the type if you have
[23:25] spare time and you're genuinely curious,
[23:27] yes. And and you have some disposable
[23:29] income, go buy a Mac Mini. Okay.
[23:31] >> Let's get into it.
[23:33] >> Okay. So,
[23:35] this is a quick presentation and I I
[23:36] think this is super useful um when
[23:39] I think
[23:40] the easier it is to set up an AI agent,
[23:42] the more
[23:44] um
[23:46] the more you are just a consumer of this
[23:49] technology. I think it is very important
[23:52] to understand, at least at a conceptual
[23:54] level, what's going on under the hood of
[23:56] an AI agent. And so,
[23:59] let's think about the goal for today
[24:01] practically, right? Build a skilled AI
[24:04] agent that can help you with work,
[24:06] right? So, in order to do this, we need
[24:09] to understand something first, right?
[24:10] What is an AI agent? So, let's dig into
[24:13] the simplest definition and this is from
[24:16] a from Claude. I loved this definition.
[24:18] An AI agent is an AI model that runs
[24:20] tools in a loop, right? Okay, what does
[24:23] this mean? Let's break this down part by
[24:25] part. An AI model, right? Like ChatGPT
[24:28] and like Claude is just input, output,
[24:31] right? I ask a question, how do I change
[24:33] a tire without using any tools or doing
[24:34] any work, right? Input and output,
[24:37] right? We've all experienced ChatGPT.
[24:40] Um I won't dive into what an LLM is, but
[24:42] everyone has had an experience with an
[24:43] AI model or at least you should have it
[24:45] by this point, right? You ask it
[24:47] something, you tell it to do something,
[24:49] text in, text out.
[24:51] Let's move to the next part, right? An
[24:53] AI agent is a model that runs tools in a
[24:55] loop. What is a tool, right? Input
[24:58] tool use, right? So, you can say, "I
[25:00] want you to search the web." And the AI
[25:03] agent can think, and then it can search
[25:04] the web. As we talked about earlier with
[25:06] tools like Claude Code, it can also
[25:08] execute code. It can browse all of the
[25:10] files. It can Remember when I said
[25:13] agents can control a
[25:15] browser earlier? AI agents can actually
[25:18] control a browser. And it can actually
[25:20] decide when to control a browser. So, if
[25:24] the if even if you don't ask for it, it
[25:26] can actually start to use tools. And so,
[25:28] this AI model can decide to use tools.
[25:31] And, you know, you can also do you can
[25:33] search YouTube. You can search Reddit.
[25:35] You can download files. You can run
[25:37] Python, right? You can send emails, send
[25:39] texts. The point of this is every day AI
[25:42] is getting better at using more and more
[25:45] tools to the point where anything you do
[25:47] on a computer, an AI agent will be able
[25:50] to do by the end of the year,
[25:52] pretty much. I I'm very confident in
[25:54] this. This is not an exaggeration.
[25:57] Okay, let's get to the final portion,
[25:58] right? An AI model
[26:00] um
[26:01] it runs tools in a loop. So, an AI agent
[26:04] is an AI model that runs tools in a
[26:06] loop. What do I mean by loop, right?
[26:08] Input
[26:09] goes into the AI model, and it can use
[26:12] tools. We already covered this. But,
[26:14] this loop right here, where it can use
[26:17] tools, the results get fed back to the
[26:19] AI model, it analyzes whatever happened
[26:22] while using those tools, and then the AI
[26:24] model can just decide to use more and
[26:26] more tools. So, when you use Claude Code
[26:28] and it runs for 30 minutes, it's often
[26:31] doing research on coding best practices,
[26:34] it's
[26:35] looking at its skills, it's using
[26:36] different skills, it's doing all of
[26:38] these things in a loop until the AI
[26:41] agent decides that it's done. The model
[26:44] will be like, "Okay, here's your final
[26:45] response." So, so let's talk about this
[26:48] in in in practical terms here, right?
[26:50] So, I'm going to say, "Build an app that
[26:52] podcaster Callum Johnson would love. The
[26:54] AI model is going to start using tools
[26:56] because the AI agent will decide when it
[26:59] wants to start using tools. So, it may
[27:00] search Callum Johnson on YouTube. I
[27:02] apologize if I misspelled your name. I
[27:05] made this super late last night. Um and
[27:07] then uh then it might search YouTube. It
[27:10] can go to your YouTube. It'll analyze
[27:11] your transcripts and then it'll get an
[27:13] idea of what types of apps you would
[27:14] want. And then, all of that data gets
[27:17] sent back to the AI model. It may think
[27:19] again and okay oh and now it's like,
[27:20] "Okay, I need to use some more tools."
[27:22] And then it may brainstorm app ideas. It
[27:24] may create a a plan for an app, send
[27:26] that information back to the AI model.
[27:28] It'll write the code. It'll build the
[27:30] app, send it back to the AI model. The
[27:32] AI model will be like, "Okay, well, we
[27:33] actually need to test the app." So, then
[27:35] it will actually spin up a browser um
[27:38] and it will actually start clicking
[27:40] around the app making sure it's good and
[27:42] it will evaluate the app and then it
[27:44] will finally send it back to the AI
[27:46] model and let's say it tested the app
[27:48] and it really liked the app. The AI
[27:50] model will be like, "Okay, that's pretty
[27:51] good." And then it'll be like, "Okay,
[27:53] now it's time to be done." And it'll
[27:57] send a message back to you and it'll
[27:59] say, "Here is your app. You can test it
[28:00] here at your app.com."
[28:02] And the point of the AI agent and what
[28:04] makes it agentic is that it will
[28:06] actually decide how long it loops for.
[28:09] You know, an automation is just input
[28:11] sets of steps, final response, right? An
[28:13] AI agent is input, model thinks, uses
[28:17] tools in a loop for as long as it wants
[28:20] to. And then once it once it thinks it's
[28:23] completed the task or sometimes once it
[28:25] decides that it's impossible to complete
[28:27] the task, it'll just be like, "Okay, I'm
[28:29] done. Here's your Here's the result.
[28:31] Here's your final response." Does that
[28:33] make sense?
[28:33] >> So, so just to jump in, Riley. So, the
[28:36] significance of this like loop that you
[28:40] explained so well is it's actually going
[28:44] to get you to like a useful output,
[28:48] right? Like cuz it's going to keep
[28:49] iterating and getting better. I think
[28:51] we've all had that experience using like
[28:53] chat GPT where, especially in the
[28:55] beginning,
[28:57] a lot of people use it like Google,
[28:59] right? You just put the question in, it
[29:01] spits out an output, and a lot of the
[29:03] times you're like, "This isn't even
[29:04] good."
[29:05] With the
[29:07] being like an AI agent, it's going to
[29:09] keep that feedback loop going until it
[29:11] actually gets to like a useful, like
[29:15] more valuable output. Is that right?
[29:18] >> Yes. And 100% and to to kind of
[29:21] piggyback on top of that,
[29:23] the more prescriptive you are in your
[29:25] original prompt, the way happier you'll
[29:28] be. So, AI models are getting much
[29:30] better at thinking
[29:31] and it much better at using tools
[29:34] to the point where it'll just kind of do
[29:36] the things you want it to do. This is a
[29:38] pretty bad prompt here. Build an app
[29:39] that podcaster Cal Johnson would love.
[29:41] You know, it's my fault if the agent
[29:43] gets back to me and I test the app and
[29:45] it's like, "Oh, this isn't that useful."
[29:46] It's like, "Well, I didn't give it any
[29:47] help." If I listed all the things that I
[29:50] wanted the app to do and then I also
[29:52] told it what style I liked or if I said,
[29:54] "Here's my website, use the style of my
[29:56] website," which AI is getting really
[29:57] good at doing,
[29:59] it would create a pro- an output that I
[30:01] that I want. And so,
[30:04] be- being a good knowledge worker or
[30:06] being a good entrepreneur at this point
[30:08] is actually just coming up with really
[30:10] good prompts or being able to describe
[30:12] exactly what you want so that as the AI
[30:15] model is going through these loops, it
[30:17] knows what the final output should look
[30:19] like. And so, that is the key. The name
[30:21] of the game is being able to clearly
[30:23] articulate what you want the AI agent to
[30:26] do because it's getting better and
[30:28] better at just doing all of the things
[30:30] that you want it to do.
[30:31] >> Yeah. You know, you mentioned this the
[30:33] significance of of prompting.
[30:36] Um and I know even there's like jobs now
[30:39] which is like the prompt engineer. Like
[30:41] your your set is actually just you're
[30:43] great at prompting AI.
[30:45] Where I wanted to go though, even in
[30:47] that example you gave of like
[30:49] business ideas or like revenue ideas for
[30:53] existing businesses,
[30:56] is that
[30:57] because I'm thinking that's like an
[30:58] incredibly
[31:00] valuable use case right there of like
[31:04] not only is it helping you brainstorm
[31:06] and kind of come up with these ideas, it
[31:07] can then like mock it up for what like a
[31:09] V1 would look like.
[31:11] Have you from your experience using it
[31:13] as someone who's actually spent time
[31:15] like hours and hours and spent money
[31:18] like using agents?
[31:20] Has that been a useful like a valuable
[31:24] use case? Like, have you seen that it's
[31:26] effective at doing that?
[31:28] >> 100% and yes, 1,000%. You know, one of
[31:32] my favorite
[31:33] uh things to do with an AI agent is to
[31:35] just kind of give it some freedom,
[31:37] right? I have some things that I want it
[31:38] to do very specifically and then I have
[31:40] some things where it's pretty open-ended
[31:42] where I will say every morning at 9:00
[31:44] a.m. I want you to analyze everything or
[31:47] analyze like maybe it's my email, my
[31:49] YouTube, um and then my like like it can
[31:52] analyze the database that we have that
[31:54] like tracks our new customers and where
[31:57] these customers are coming from. I want
[31:59] you to analyze it and I want you to come
[32:00] up with a useful insight based on all of
[32:03] this data. And AI is really good at
[32:05] analyzing data and then one day it'll
[32:07] just surprise you. Be like, "Hey, I
[32:08] think you should do this because of
[32:10] this, this, and this, and this. Here
[32:11] someone, you know, eight people
[32:13] commented on your videos that they want
[32:15] this feature. Like, you should build
[32:16] it." And then since it has access to
[32:18] your uh linear, it's like it'll be like,
[32:21] "Oh, Saul on your team isn't that busy
[32:23] and based on his previous tasks, I think
[32:24] he could create this." You know, like
[32:26] it'll it
[32:28] you know, the IQ of an AI agent is
[32:30] getting insanely high, you know, like
[32:32] Opus 4.5 I think it's like 130 that that
[32:35] of a IQ of 130. If you give it all of
[32:37] your data and you say analyze it, act as
[32:40] a
[32:41] um act as a
[32:43] chief operating officer and just analyze
[32:46] it and come up with one useful
[32:48] um idea every single day and prepare it
[32:50] in a concise one-page report and include
[32:53] hyperlinks to like relevant links, that
[32:55] can be incredibly useful. And so, you
[32:58] know, like while I'm drinking my coffee
[33:00] in the morning, I can just read this
[33:01] report that an AI agent prepares. You
[33:03] don't have to use it, right? Some of the
[33:05] ideas might You might be like, oh, I
[33:06] already had that idea or but like
[33:08] sometimes it'll be incredibly
[33:10] insightful. And so, yeah, to answer your
[33:12] question, like yes, it is incredibly
[33:15] useful to create workflows like this.
[33:17] >> Yeah. You know, as you were speaking,
[33:18] I'm like coming up with And I remember
[33:20] this happened the last time I was on
[33:22] >> I think we'll get there. Like I In this
[33:24] video, we're going you know, the next
[33:26] step of this is we'll create an agent
[33:28] and we we'll create one together. Like
[33:29] we can come up with an idea and I'll
[33:31] show you exactly how to set up a cron
[33:33] job, which is just an automation, and
[33:35] then we can test it and we can actually
[33:36] see if it's valuable or not. So,
[33:39] yeah, yeah, I think I think that'll be
[33:41] really fun. Um here, can I just finish
[33:43] this up real quick? So, I just wanted to
[33:44] get to one other point here
[33:46] um
[33:47] around AI agents that run tools in a
[33:49] loop, right? We covered AI model, we
[33:50] covered tools, we covered loop, right?
[33:52] And I just kind of wanted to tactically
[33:54] show you this is from
[33:56] 2025. You know, I'm pretty sure it's up
[33:58] to 7 hours now. But look at this curve
[34:01] of of this is the time horizon of
[34:03] software engineering tasks
[34:05] um that LLMs can complete 50% of the
[34:08] time. So, this is So, GPT-5, when it
[34:11] came out in 2025, could do tasks that
[34:14] were over 2 hours. We're up to 7 hours
[34:17] now in 2026, and this is only going to
[34:19] increase. And you know, and by the end
[34:21] of the year, I said this earlier, we'll
[34:22] be measuring AI agents in days, not
[34:26] hours. We'll be like, okay, that you
[34:27] know, AI last month it could work for 4
[34:30] days, now it can work for 2 weeks. Um
[34:33] and that's coming. And so, if we get
[34:35] back to our original goal, right? Of we
[34:37] want to build a skilled AI agent that
[34:39] can help you with work. In order for
[34:41] agents to run most tools, you need to
[34:43] give the agent access to a computer. You
[34:46] brought up a Mac Mini. I said you
[34:47] actually don't necessarily need a Mac
[34:49] Mini. You could use a MacBook Pro or you
[34:51] can use a sandbox.
[34:52] Um
[34:53] just a hosting service that runs these
[34:56] agents in the cloud, right? Um and for
[34:58] coding, you can use Cloud Code. For co-
[35:01] you for work, you can use Co-work. You
[35:02] can use Open Claw. I didn't include
[35:04] Perplexity Computer, Manifold, and there
[35:07] are some others and there will be many
[35:08] more.
[35:09] Um for the sake of this, I like to use
[35:11] Open Claw. So, I'm going to show you how
[35:13] to get set up on Open Claw. Um do you do
[35:17] you want to hop into this or do you do
[35:18] you have any questions? Okay.
[35:19] >> I'm excited. I'm like, let's do it.
[35:22] >> Yeah, let's create it. So, for this, I'm
[35:25] going to use So, you can use Hostinger.
[35:28] You can use Chorus, which is what I use,
[35:31] chorus.com. You can use
[35:34] Um you can create a VPS directly on AWS.
[35:37] There's many services that you can use
[35:39] that allow you to spin up an agent on um
[35:43] on
[35:45] um on the web. And so,
[35:48] this is a brand new agent that I created
[35:50] right before you and I'll show you the
[35:51] process of creating an agent. Um you can
[35:53] just literally create new agent and you
[35:55] can just click like salesperson. And
[35:57] then, you can go through and just
[35:58] connect it to all of your things, right?
[36:01] And you go through this process. When I
[36:02] hit done, it does take 2 minutes to
[36:05] because it needs to spin up a virtual
[36:07] computer and I don't really want to wait
[36:09] for that. So, I've already gone through
[36:10] this and I have it set up.
[36:13] Um let me see here. Yeah. So, this is
[36:16] just this Riley's Open Claw agent. I
[36:18] haven't edited or added any integrations
[36:21] or skills yet. So, after you go through
[36:23] the process, if you don't set up
[36:24] everything, this is what you get. You
[36:26] get an AI agent with a computer. And so,
[36:29] all this is is just like an AI chat um
[36:32] running Open Claw. And here we see all
[36:35] of the files that Open Claw starts out
[36:37] with. What separates
[36:39] >> Just to be clear, Riley, so just uh so
[36:42] people can follow along. This is
[36:46] This is almost like the sandbox. Like
[36:48] this is the virtual computers like
[36:51] running it. That's why you're using a
[36:52] hosting uh or Chorus. Like explain that
[36:56] part of it.
[36:57] >> Yes. When you Let's say you you bought
[37:01] into all the hype that all the
[37:02] influencers were talking about and you
[37:04] went out and bought a Mac mini. And then
[37:06] you brought it home and you set it up
[37:07] and then you went to Open Claw's website
[37:10] and you installed Open Claw. What it
[37:11] would do is it would create a directory
[37:13] on your computer that looks identical to
[37:15] this.
[37:16] Um and you would have these certain
[37:19] files
[37:20] uh that make up your agent's personality
[37:24] and you would be it's pretty easy to set
[37:25] up skills. But, the reason like some of
[37:28] these hosting services are better is
[37:30] like it's super easy to just set up
[37:31] skills. Like you can go to the skills
[37:32] tab and you can go to like the
[37:34] marketplace and you And so, like these
[37:36] skills
[37:38] are basically packaged abilities for
[37:40] your AI agent that you can just enable.
[37:43] And your AI agent will actually is like
[37:46] consciously aware of what skills are
[37:47] available. Like it can go check popular
[37:49] marketplaces of skills and it can
[37:51] install them just by Like you can just
[37:53] be like, "Hey, do you want me to do it
[37:54] to install the 11 Labs skill so I can
[37:56] speak to you?" And then you can just be
[37:58] like, "Yes." And then it will just do
[37:59] this automatically. Um And so, this This
[38:03] is just a very easy way to spin up an
[38:05] Open Claw and then see all of its files
[38:08] and uh basically easily connect it to
[38:12] all of your stuff. You know, and so
[38:15] that's what this is. This is a computer
[38:17] that runs 24/7 unless you delete your
[38:19] agent. I'm pretty sure, yeah. You can
[38:21] like delete your agent right here. And
[38:23] um unless you delete your agent, this
[38:25] will run 24/7 and it will be as if you
[38:28] have a physical computer running. It's
[38:30] just running in the cloud.
[38:31] >> Okay, su- super cool. And And you know
[38:33] what? I love the way that you explained
[38:35] it where the skills, it's almost like
[38:38] you're giving your agent abilities.
[38:42] And so the question that instantly comes
[38:43] to mind for me,
[38:45] is there such thing as giving my AI
[38:47] agent too many skills? Like too many
[38:50] abilities. Is that a problem that
[38:54] people, especially in the beginning,
[38:56] kind of stumble into where like it can
[38:59] it can almost do too much or too little?
[39:02] Share Share that perspective.
[39:04] >> Yes. So, this is a great question and
[39:07] I'm glad you asked it because people's
[39:09] first instinct when they set up Open
[39:10] Claw just should just They just So
[39:12] there's like, "What skills can I add?"
[39:13] And they just start adding skills like
[39:15] willy-nilly going crazy.
[39:16] >> That's what I would do. Yeah.
[39:17] >> think Right. No, no. And I That's what I
[39:20] did first, too. I did it twice. I made
[39:21] the same mistake twice and I had to cut
[39:23] the entire agent because it had like 80
[39:25] skills. And um
[39:27] if you think about it, just think about
[39:28] an employee, right? If you were to hire
[39:30] an employee and you know, I'm way more
[39:33] comfortable hiring an employee that says
[39:34] they're really good at two things and
[39:36] they can achieve a specific goal, right?
[39:38] And And I'm just like, "Yes, I want to
[39:39] hire this employee because they can
[39:40] predictably help me in X, Y, and Z
[39:43] because they're good at A, B, or C."
[39:46] Right? As soon as you get that employee
[39:47] who's like, "Oh, yeah, I can do this. I
[39:48] can do this. I can do this. I can do
[39:50] this. Oh, we can help you with this."
[39:51] Right? It gets scattered. And then the
[39:53] agent, because all of that information
[39:55] is stored on their computer. They're
[39:56] just files. You know, we can actually
[39:58] click into the skills tab right here.
[40:00] Like you can go into the skills tab.
[40:02] This is just a representation of files
[40:04] that are running on your computer.
[40:05] Skills are just markdown files in the
[40:08] skills folder. This is the same for
[40:10] Claude. This is the same for Open Claw.
[40:12] Here, right? We see home, right? This is
[40:14] on your computer. You see skills. And so
[40:16] as soon as you go and download like 30
[40:18] skills to your computer, right now your
[40:22] agent, anytime your agent wants to do a
[40:24] task, it analyzes what skills it has
[40:26] access to and it only uses the relevant
[40:28] one when it needs to. If you have too
[40:30] many skills, it's going to use the wrong
[40:33] skill
[40:34] uh at a higher frequency. And so, I
[40:36] found a sweet spot between like 7 and
[40:38] and up to 15 to 20 skills. Anything
[40:42] above 20 skills, I think you get a steep
[40:43] drop-off and it's not nearly as good.
[40:47] And so, that's why I actually have
[40:49] multiple agents. And so, here on the
[40:51] left here, these Zompa, I don't know
[40:53] where I came up with the name Zompa.
[40:55] This is my personal agent that I use
[40:56] through iMessage. And so, the coolest
[40:59] thing that OpenClaw did is if you go
[41:01] into connections and go into
[41:02] communications, you can see here that
[41:04] you can just text it. Um or you can go
[41:06] into Telegram and you can add it to
[41:08] Telegram. Uh or you can go into Slack,
[41:10] you can add it to your Slack. And so,
[41:12] that's what's happening right now is
[41:14] people are putting these useful agents
[41:16] that control a computer into where you
[41:19] already message other humans. And so,
[41:20] now you can just like add them to group
[41:22] chats. You know, I
[41:24] that technology hasn't been fully
[41:25] figured out. Like, it's not perfect yet
[41:27] for creating group chats, but that's
[41:29] coming.
[41:29] >> To your point, it's like a a super
[41:31] powerful
[41:32] uh like assistant in that way.
[41:35] >> Yes.
[41:36] >> And then even
[41:36] >> Okay, let me
[41:37] >> I'm and even just Well, go ahead.
[41:40] >> No, I would love to kind of You were
[41:42] talking about prompting earlier and I
[41:43] would love to kind of talk about what's
[41:45] next in terms of like prompting and it
[41:48] has to do with these files right here.
[41:51] Um
[41:52] and so,
[41:53] like this is really cool. So, I have
[41:55] You've never used OpenClaw, right?
[41:58] So, what separates OpenClaw from
[42:01] Let's close this out. Let's close out
[42:03] everything that's not relevant, right?
[42:04] All we have here is just an AI chat that
[42:07] has access to files, right? It has
[42:09] access to these files. And so, it says,
[42:10] "Hey Riley, uh what do you want uh this
[42:13] thing to help you with, right?" It'll
[42:15] just say something at the beginning. And
[42:16] I'm going to say, "Hey,
[42:18] um please
[42:20] um
[42:21] Like you can actually start off So,
[42:23] let's think about what we want this
[42:24] agent to do. Hey, I want Hey, your new
[42:28] name is
[42:31] um
[42:32] CJ.
[42:34] And I want you to be uh to help me
[42:41] grow my YouTube.
[42:44] And so, let me show you something here.
[42:45] So, when you message this Open Claw
[42:47] agent
[42:48] So, I have this open over here.
[42:51] Oh. Oh, yeah. So, I guess the onboarding
[42:54] changed for this. So, here it just
[42:56] suggested skills that we can use. So,
[42:58] I'm going to do YouTube search, YouTube
[43:01] competitor analyst, trends spotter. So,
[43:05] you can see here Open Claw just kind of
[43:06] like came up with skills that would be
[43:08] useful for me.
[43:09] And now it's setting them up, which is
[43:11] pretty cool.
[43:12] >> So, it's like walking you through It's
[43:14] almost like handhold like step-by-step
[43:17] through the process of getting you
[43:19] onboarded.
[43:20] >> Yeah, and it's it's um
[43:23] Open Claw did this first, and that's why
[43:25] it's was the fastest-growing software
[43:28] ever created. Oh, service is overloaded.
[43:30] I guess it's getting a lot of usage. Um
[43:33] here, we can just create a new chat
[43:34] here. Let's create a new chat. We don't
[43:35] want to be on the onboarding. I want you
[43:38] to um
[43:40] edit your um
[43:43] your soul.md
[43:46] file such that
[43:49] um soul.md file and other necessary
[43:52] files. All right, I like to say this at
[43:54] the beginning, files. So that you are
[43:58] obsessed
[44:00] with helping me grow my YouTube channel.
[44:05] Right? Here, it's like we're telling
[44:07] >> Step one is like you're just stating
[44:09] your goal. Just plainly simply stating
[44:12] your goal.
[44:13] >> 100% The first thing that you should do
[44:15] with your agent when you start off,
[44:17] right? So, right now it's like now let
[44:19] me update soul.md and identity file.
[44:22] Let's take a look at this beforehand.
[44:23] Hopefully, it doesn't change first. This
[44:26] is open claw, right? Open claw
[44:28] Open claw created these files called
[44:31] soul.md. So, as you use your agent, it
[44:34] will automatically update these files,
[44:36] right? It'll update the soul.md file
[44:39] and the soul.md file. So, we can
[44:41] actually see here. We can click on
[44:43] identity.
[44:44] Your name is Olga. And I can say no,
[44:47] make your name CJ, right? You can pick
[44:50] its name. And this is useful when you
[44:53] have multiple agents and it starts
[44:54] messaging you on texting here.
[44:56] But you can see here it's actually like
[44:57] making changes to the different files.
[44:59] And you can see like here's the
[45:01] identity,
[45:02] the user.md.
[45:04] It knows my name is Riley Brown. And
[45:06] when I first set it up, it sets up with
[45:08] my email. And so, it actually went in to
[45:11] my integrations. Actually learned about
[45:13] me just based on my email.
[45:15] Notes it's So, I can say, no, make your
[45:17] name CJ. My account is and then we can
[45:21] actually just go to YouTube real quick
[45:24] and we can snag the URL or we can tell
[45:26] it to look it up, but I can just go
[45:29] Riley Brown and uh
[45:31] >> As you do it, what comes to mind? It is
[45:33] like having It's like onboarding an
[45:36] employee. I think if you had like an
[45:37] employee joining your organization or
[45:40] your team at work, like all of the
[45:42] things that you would have to share so
[45:43] that they could then go and do their
[45:45] job.
[45:45] >> Yes. And
[45:47] And you're always doing So, when you
[45:49] asked earlier you said like, is prompt
[45:51] engineering a real
[45:53] profession? And I would say no. You
[45:54] know, and um
[45:56] What it's uh someone came up, I think it
[45:58] was Andre Karpathy. He started talking
[46:00] about how it's context engineering. It's
[46:03] not prompt engineering, it's context
[46:04] engineering. And so your goal whenever
[46:06] you're onboarding an agent or working
[46:08] with an agent, you want it to be aware
[46:10] of context. You want to guide it to the
[46:11] right context. And let me show you what
[46:13] I mean by this.
[46:14] I think, you know, we're onboarding this
[46:17] guy to help me grow my YouTube, right?
[46:18] I've grown from 40K subs to 200K subs uh
[46:21] last year. This year I want to go I want
[46:23] to get to 500K, let's say. Um
[46:26] if I'm onboarding this person, I want
[46:27] them to understand my goals, right? In
[46:29] in order for it to suggest good videos
[46:31] to make, it should know my last 10
[46:33] videos. It should know the transcript.
[46:36] It should know these things. Well, guess
[46:37] what? All of that is built into Open
[46:39] Claw like Open Claw. And so when you use
[46:41] like Chorus on Chorus, they have like a
[46:43] skill here. This super data allows you
[46:45] to just find any link on YouTube and
[46:47] immediately convert it into a
[46:49] transcript. And so I can say, "Please
[46:50] look at my channel. Please look at my
[46:53] last uh 10 videos
[46:57] and summarize
[46:59] my
[47:01] um
[47:01] my interests,
[47:04] my teaching style,
[47:06] and my hooks. You know, this will take a
[47:09] little bit longer, right? Cuz it has to
[47:11] do all of this these tasks. And like I
[47:14] said, you know, if we go back to um if I
[47:16] go back to this slide right here, right?
[47:20] You know, soon it's going to be hours
[47:23] that this agent can just work for you. I
[47:24] mean, this will take 10 minutes, but
[47:26] like soon you'll be able to give it way
[47:28] longer tasks. Um okay, summarize my
[47:31] interests, teaching style, and anything
[47:34] that would be relevant uh to you, uh
[47:38] relevant to you
[47:40] um for starting as my assistant.
[47:45] Right?
[47:45] >> So one one thing one thing that
[47:47] immediately comes to mind cuz I
[47:48] literally Riley, I literally did this I
[47:50] want to say 2 weeks ago. Um cuz I was
[47:53] trying to get set up on Claude. And I
[47:56] wanted to create I wanted to create like
[47:57] a project which had this context on me.
[48:00] And so I had to go into my YouTube
[48:02] channel, get the most popular videos,
[48:04] figure out a way to get the transcript.
[48:06] I then had to upload the transcripts
[48:08] like PDFs one by one into Claude's like
[48:13] back end so it could it had this
[48:15] context. If I'm not mistaken, you just
[48:19] you put super data as one of the skills
[48:23] and then you just put the link to your
[48:25] channel
[48:26] and it's able to go and do that work
[48:28] itself.
[48:30] >> Yes, it would have found it regardless.
[48:31] I could have said, "Hey, my name is
[48:32] Riley Brown. I make content on AI." Like
[48:34] it would have gone and searched. It
[48:36] would have figured out a way to find
[48:37] that information.
[48:39] Um
[48:40] Yeah, like look, it's like saying like
[48:42] now let me pull transcripts from a few
[48:44] ones and I'll analyze it. And you see
[48:45] it's just kind of going at its own
[48:46] thing. Like it's still working. And
[48:48] what's cool about this
[48:49] um
[48:50] what's cool about this is soon, you know
[48:52] how like we're adding skills and then
[48:54] using it? Soon, these AI models, you'll
[48:56] give it a long task and in the same
[48:59] response, it'll add the skills it needs
[49:02] and then it will use them in the same
[49:04] like run. So it'll it'll get a skill and
[49:06] use it in the same run. So like you
[49:08] might give it a long task and you might
[49:10] come back to like a fully completed
[49:13] document of whatever you're trying to do
[49:14] and you'll find out that it gained three
[49:16] skills in that process.
[49:18] Um which is which is pretty interesting.
[49:20] >> So it's almost like it's it's over time,
[49:23] it's just going to get more and more
[49:24] autonomous. And I and I think about what
[49:27] you showed in the beginning in that
[49:28] diagram, right? Where you have like the
[49:30] input. It feels like a year ago when I
[49:33] was using Chat GPT or Claude, I had to
[49:36] put so much input and so much context
[49:39] and like give it all of these things so
[49:41] it could give me the output
[49:43] I wanted. It feels like increasingly
[49:45] over time, we're getting to the point
[49:47] where you just state the outcome that
[49:50] you want, which is like I want to grow
[49:52] my YouTube, it will make sure that it
[49:55] has everything it needs and the context
[49:57] and transcripts and skills in order to
[50:00] achieve that outcome.
[50:02] >> Yes, and your job now, while it's still
[50:06] not perfect, right? It's not superhuman
[50:07] intelligence, it doesn't just have
[50:09] control over everything, is you just
[50:11] want to be able to monitor its outputs
[50:12] and be like, okay, what context would
[50:15] have been more useful? And what you can
[50:17] do the reason so these skills are almost
[50:19] like recipes or you can think of them as
[50:21] just standard operating procedures. You
[50:23] know, I have an SOP for everything in my
[50:25] content pipeline, like the way my editor
[50:27] edits my videos,
[50:28] there's a standard operating procedure
[50:30] and you can figure out where the agent
[50:33] commonly makes mistakes and then you can
[50:35] go in and just change the skill file.
[50:36] Yes, I can just automatically install
[50:38] these, but I can also come in here and
[50:40] edit the skill directly, right? You can
[50:42] you can literally edit the skill. And so
[50:45] you might just put you know, you can
[50:47] create your own skill and make it a um
[50:50] and make it um a standard operating
[50:53] procedure and then when you notice the
[50:54] AI make mistakes, you just include it in
[50:56] the skill so that it doesn't make a
[50:58] mistake in the future. And so the game
[51:00] is not only stating the outcome, but
[51:02] it's also evaluating the outcome based
[51:04] on what you wanted, right? Which is
[51:05] called evals. You know, a lot of there's
[51:07] a lot of companies that
[51:09] um all they do is they focus on evals,
[51:11] like evaluating how good your output is
[51:14] and depending on that, you're just kind
[51:15] of changing the prompting or the skill.
[51:17] And so
[51:18] or the skill file. So that's the game.
[51:20] It's like, what do you want? What did
[51:22] you get? How do you improve the the the
[51:25] prompt or or the system so that it gets
[51:28] closer to what you want?
[51:30] >> Mhm.
[51:31] You know, as it's as it's working
[51:33] through this because I know that
[51:35] >> done, I believe.
[51:36] >> Okay.
[51:37] >> Um
[51:39] Yeah, so here it just it created So
[51:41] again, this is this is all really good.
[51:43] So as you can see here, you have the
[51:44] home file and if we click on this, you
[51:46] see that the sub file here is memory,
[51:49] and then it just logged this on 4 2. So,
[51:53] when the agent deems that it is
[51:55] important to log a memory, it just logs
[51:57] the memory, right? We're talking to this
[51:59] agent, and then it just creates a file
[52:01] in the computer that it's running in,
[52:03] right? And you can see here, it's like
[52:04] session start, channel analytics, um and
[52:07] it's like hook patterns, bold claim
[52:09] proof pattern, cuz it analyzed all of my
[52:11] um
[52:12] my It analyzed 10 transcripts. It pulled
[52:15] 10 transcripts, these videos right here,
[52:17] and these are all of my recent um uh
[52:20] videos. And then I can say, "Please make
[52:23] this a website uh with a public link I
[52:27] can send public link I can send uh to
[52:32] friends
[52:34] um
[52:35] with
[52:37] um
[52:38] with
[52:41] uh hooks written in full of my top five
[52:47] videos from this. Create a report."
[52:52] >> You
[52:52] >> And so,
[52:53] >> as it's doing that, you cuz you
[52:54] mentioned memory, like the significance
[52:56] of memory.
[52:58] Can you explain, first of all, what
[53:00] memory is, and then as we get further
[53:03] along the process, like we get deeper in
[53:05] the process,
[53:06] what is the impact, like what is the
[53:10] the the impact, to be honest, of it
[53:13] having memory?
[53:14] >> So, what we talked about why Open Claw
[53:17] was super useful, and the first thing
[53:19] was the fact that it made it super easy
[53:20] to for you to connect it to
[53:23] um like Intercom and and Telegram, etc.,
[53:26] and all of your emails, and basically
[53:28] everything that you already use. And
[53:29] then they made it super easy to install
[53:31] skills.
[53:32] The third thing that it did is they had
[53:34] it automatically just log um
[53:38] log memory files. And so, if you go into
[53:41] the files, you see a file, like we see
[53:42] here, right? The top one is just memory.
[53:44] It has a memory folder. When it does
[53:48] something important that like maybe it's
[53:50] the like based on its system prompt of
[53:52] OpenClaw, it'll just log memories. The
[53:55] reason that's important is you may ask
[53:58] your agent to do a task 3 months from
[54:00] now. And what it can do is it can just
[54:03] search the memory, right? It can just go
[54:05] back to the memory files and it can
[54:07] actually figure out what you've done
[54:08] previously. And that can actually help
[54:10] it, right? It'll just log its own
[54:12] insights. Or, you know, and and part of
[54:15] being a good prompter or being good with
[54:17] AI agents right now is um
[54:20] is basically
[54:22] if something is important, you can say
[54:24] like I want you to always remember this.
[54:26] Please always remember this. And so, it
[54:28] can it'll either log it in the skills
[54:30] file if it if you have a specific skill
[54:32] it can attach a memory to, it'll do
[54:34] that, or it'll just log these memory
[54:36] files, which it will check um often. And
[54:40] as AI models get better, it'll actually
[54:42] get better and better at using memory.
[54:44] And so,
[54:45] an AI model that just kind of remembers
[54:47] things is way more useful than one that
[54:49] doesn't. Does that make sense?
[54:50] >> Yeah. And it's also
[54:53] also what comes to mind for me,
[54:55] as you continue to use the agent, it's
[54:58] going to get more useful for your
[55:00] company cuz it's going to remember
[55:03] certain pieces of data and information
[55:05] and then be able to like hearken on that
[55:09] in the future.
[55:11] >> Yes, 100%. Here, it's done with the
[55:13] website. So, we can click on this. Boom.
[55:16] So, look at this. So, I told it to
[55:18] create this report, right? And, you
[55:20] know, it's
[55:21] it's just top five videos with opening
[55:23] hooks. So, I gave OpenA OpenClaw blender
[55:26] skills and look, it probably
[55:27] automatically embedded the link into it.
[55:29] Yep. There you go. So, like you can just
[55:32] click on it. You have the links. Here it
[55:34] has the exact hook that I used. Today
[55:36] we're going to dive into the most viral
[55:37] AI tools. Open claw. Yep, this is the
[55:38] exact hook, right? It has the amount of
[55:40] views, the likes, 15 minutes.
[55:43] One integration I'm I'm working on, it's
[55:44] actually hard to set up, is your YouTube
[55:47] analytics platform, which actually has
[55:49] all of the retention data. So, you know
[55:51] how like Mr. Beast relentlessly analyzes
[55:54] your retention, his retention, and sees
[55:56] where people drop off. I'm sure you do
[55:58] the same. Like, your intros are for your
[55:59] podcast are excellent, and you probably
[56:01] found that like when you do an intro
[56:03] like this, you have better retention.
[56:05] When you have an intro like this, you
[56:06] might get worse retention. Your AI agent
[56:08] will actually be able to just analyze
[56:10] all of that data. And so, if you give it
[56:12] a long enough time horizon to work, you
[56:14] know, and maybe you want to give it 3
[56:15] hours to um and we'll talk about
[56:18] creating these super long workflows in a
[56:20] bit. Um you can have it analyze. Be
[56:22] like, "Okay, find Look at all of my
[56:24] videos I made in the last 6 months.
[56:26] Analyze the hook and the retention, and
[56:28] figure out where like what videos were
[56:31] the best, what videos were the worst."
[56:33] And it will give you a great It'll
[56:35] surprise you. It'll give you a really
[56:36] great response. The problem is is that
[56:38] integration right now is a little bit
[56:41] hard to set up, but um
[56:43] uh I figured out how to make it easier,
[56:45] so
[56:46] I'll be talking about that in the
[56:47] future. Um but like here, hook patterns
[56:49] that work. Bold claim plus live proof.
[56:52] Everyone's doing it wrong. Replace
[56:53] expensive thing. I built X without Y. Um
[56:58] And then here it has teaching style, and
[57:00] like this is just a website. And like I
[57:01] could literally send you this this link
[57:03] right here, and it will just work. Um
[57:06] and so, these get stored I think these
[57:08] get stored HTML file. Um yeah, so this
[57:11] is just in your public folder. And
[57:13] again, all of your apps will just go in
[57:14] this public folder, and you can open
[57:16] them up at any time. You just click on
[57:17] it. Um and so,
[57:20] the same way you can have
[57:22] uh little apps on a real your real
[57:25] computer, right? I go to my dock, I have
[57:26] all these apps running. I can have
[57:28] little apps in my virtual computer that
[57:29] my agent can create, which is super fun.
[57:32] And this is just a single file, but this
[57:34] is super useful.
[57:35] >> Yeah. You know, I'm I'm curious now,
[57:37] Riley, because um
[57:39] and obviously you can share as much as
[57:40] you want, but
[57:42] looking at your channel in the last 5
[57:46] I'd say the last 5 to 6 weeks, there's
[57:49] been like a spike even in the
[57:51] performance of your videos.
[57:54] How much have you been using OpenClaw
[57:57] and like insights from these agents
[58:00] to actually help inform what videos
[58:03] you're creating and I don't know, the
[58:05] scripting and the thumbnails and the
[58:08] packaging. Like how much is
[58:10] OpenClaw responsible for that
[58:13] performance?
[58:15] >> Um
[58:16] how much is it responsible for the
[58:18] performance? You know, I maybe it's a
[58:20] 10% increase in um quality of video. So,
[58:24] I do have OpenClaw writes on my hooks.
[58:26] The way that I I think of my video, I
[58:28] think of my content, my content's pretty
[58:30] raw. I like to talk, I like to go on
[58:32] tangents, I like to show my screen. Uh
[58:34] so, I decide what the most valuable
[58:36] thing I can share is, but actually I get
[58:38] a lot of those insights from my YouTube
[58:40] comment analyzer. I actually just
[58:42] realized it's probably more important
[58:43] than I think, cuz like you said or um
[58:46] like I like I said earlier, I do have
[58:48] that automation that goes every week
[58:50] that tells me exactly what people are
[58:51] asking questions on, and that's actually
[58:53] how I came up with um this specialized
[58:56] agents. Because people were like, "Oh, I
[58:58] added too many skills. Like how do I How
[59:00] do I manage the amount of skills?" And
[59:02] so, I made a video dedicated to that.
[59:04] Boom, 100K views. So, actually more.
[59:06] It's It really helps my decision-making,
[59:08] but what I'll do is I'll outline the
[59:10] whole video that I'm going to make, and
[59:11] then I'll film the whole video, and then
[59:13] I'll film the intro, and I'll you And
[59:15] now I have a workflow that like I feed
[59:17] it the transcript of my of the like the
[59:19] body of like, you know, the 40 minutes
[59:21] that I filmed for my video, and I say,
[59:23] "Come up with an intro." It'll analyze
[59:25] all of my best performing intros, and
[59:27] then it will just write the intro. Maybe
[59:28] I'll make a few tweaks, and then I'll
[59:30] just film it on my teleprompter. And I
[59:32] I'm actually talking to you using a
[59:34] teleprompter right now. And so, yeah,
[59:37] maybe it's actually like
[59:38] pretty important, and I didn't even
[59:40] realize it.
[59:40] >> Um that's fascinating.
[59:42] >> Yeah.
[59:43] >> That's fascinating.
[59:44] >> Um
[59:45] Yeah, what else do I want to show? Okay,
[59:47] um if you don't mind, I would love to
[59:48] kind of go over cron jobs, because this
[59:50] is part This is, I think, one of the
[59:52] most important things. Um
[59:54] And also, actually, first, let me show
[59:57] you something. So, what we can do here
[59:59] is, let's say you use Telegram. So, this
[01:00:01] is Telegram right here.
[01:00:04] And um
[01:00:05] all you do,
[01:00:07] uh if you want to like add a new
[01:00:08] account,
[01:00:09] right? This process is going to get a
[01:00:12] lot easier, I think. I think Telegram's
[01:00:14] just going to kind of add it. Right?
[01:00:15] They have this weird feature where you
[01:00:17] have to go to BotFather, and you have to
[01:00:18] hit like new bot, and then it'll um and
[01:00:21] then you name it, right? I'll just say
[01:00:24] um you don't actually like
[01:00:27] you can just say like uh CJ, and then it
[01:00:29] You have to give it a name for a bot,
[01:00:30] like I'll call this like YouTube
[01:00:32] assistant guy
[01:00:35] um bot.
[01:00:37] And so, that's just an example of a way
[01:00:39] to connect it to Telegram. And so, we
[01:00:40] can just immediately connect it to
[01:00:41] Telegram. The agent will then say,
[01:00:43] "Okay, let's get started," and it will
[01:00:45] actually just configure it. You can say,
[01:00:47] "Hi, are you uh live?"
[01:00:49] Um See? CJ's already typing, right? And
[01:00:54] um you say, "Yep, alive and beautiful.
[01:00:56] What's up, Riley?" It already knows me.
[01:00:57] So, this is a channel that I already
[01:00:59] use. The agent is alive in Telegram.
[01:01:01] >> So, if you if you if you had it
[01:01:03] installed on your phone, like Telegram,
[01:01:05] and you go into the app, you could see
[01:01:07] conversations, or even in like a group
[01:01:10] chat. You would be able to see this
[01:01:12] agent that you've set up, like its
[01:01:15] responses, like a conversation.
[01:01:17] >> Yeah, yeah, yeah. It This is my agent.
[01:01:19] And so, I can say like, "Hey, um
[01:01:22] like I could
[01:01:23] and so like we could whatever we do
[01:01:25] here, right? We can go to skills. I
[01:01:26] actually don't know if we connected.
[01:01:28] Yeah, I didn't connect my um what's
[01:01:31] really important for me is connecting
[01:01:33] notion, right? Look at how easy setting
[01:01:35] up notion is.
[01:01:36] Boom. This is my studio team. Boom,
[01:01:38] done, right? If we create a new chat and
[01:01:41] there's skill available. So here based
[01:01:45] on the connections, it has recommended
[01:01:47] skills for those connections. And so you
[01:01:49] can just hit install, right? And so now
[01:01:52] it installed the
[01:01:54] Here it installed the Google skill. Um
[01:01:58] here we have the notion skill. So we can
[01:02:02] just immediately use notion. And so if
[01:02:04] we go to the main chat actually let's go
[01:02:06] to this chat right here. Um
[01:02:09] like
[01:02:10] one thing that you can do if you want to
[01:02:12] like reset your chat. This is when you
[01:02:14] get kind of deeper into kind of the
[01:02:15] slash commands. You can like reset your
[01:02:17] agent, right? There's a lot of things
[01:02:18] that you can learn. Um now what we can
[01:02:21] do is we can say
[01:02:23] um I'm going to say first of all, do you
[01:02:25] have access to notion? I always like to
[01:02:29] check. Sometimes it takes a little bit
[01:02:30] longer to like set up the skill. Uh I
[01:02:33] can say I just added the integration
[01:02:36] and skill.
[01:02:38] >> Can you share like
[01:02:40] what are examples of how you're using it
[01:02:42] with notion? Like what's the what's the
[01:02:43] value of that integration?
[01:02:46] >> So
[01:02:47] again, my notion
[01:02:49] um we can bring up notion real quick. So
[01:02:53] um
[01:02:55] here
[01:02:56] >> It's like your content database.
[01:02:58] >> Yeah, we have a content database, right?
[01:03:00] On my table, if it shows the ones that
[01:03:02] are like it's just all of my content
[01:03:03] that I've ever created. And I even in
[01:03:06] here I have all the comments. So first
[01:03:08] of all, it can actually just like add
[01:03:10] stuff to notion here. So let's say hmm,
[01:03:12] not seeing a notion skill. Interesting.
[01:03:15] Um
[01:03:18] um
[01:03:22] added
[01:03:24] Sometimes it takes a little bit to
[01:03:25] install. I just added Notion um skills.
[01:03:28] Maybe we can search Notion Marketplace.
[01:03:30] >> It's actually it's helpful for people to
[01:03:32] see this.
[01:03:34] >> Yes.
[01:03:35] Um and and you just want to make sure
[01:03:37] that it just has access to stuff. Like
[01:03:38] when you add Notion, you want to make
[01:03:39] sure that it has access to Notion. I
[01:03:41] just added Notion.
[01:03:42] Um
[01:03:43] >> And again, it's like strikingly similar
[01:03:46] to like having an employee. You know,
[01:03:49] you think you've given them access to a
[01:03:51] Google Doc or an email or like an
[01:03:53] account that they would need, but then
[01:03:54] they don't have it. And so you have to
[01:03:56] Like that process actually feels very
[01:03:58] similar of like onboarding an employee.
[01:04:01] >> Yes, exactly. It's It's very similar.
[01:04:03] And so I'm going to say, "Can you please
[01:04:05] look at" and we can go to You can just
[01:04:07] test this out. So this is the Riley
[01:04:09] content database. I could give it this
[01:04:10] exact link, but it will just go find it.
[01:04:12] Riley content database, tell me what
[01:04:15] videos I've used recently. Um
[01:04:20] or I've used or and then I can just do I
[01:04:22] I've
[01:04:23] made. I don't know why I said used.
[01:04:27] So in here. So
[01:04:30] what all software companies are doing
[01:04:31] right now and what Notion is Notion's
[01:04:33] becoming an AI-first company. And that
[01:04:34] doesn't mean like they're adding AI to
[01:04:36] their app necessarily, which they are,
[01:04:38] but they're also making it readable by
[01:04:39] any AI agent. So take a look at this,
[01:04:42] right?
[01:04:43] Um
[01:04:44] you know, here we have like all of the
[01:04:46] content that I've done, right? This is
[01:04:47] already useful. And I can say
[01:04:51] um please
[01:04:52] So I could say ple- um I have an idea uh
[01:04:56] for a video. I'm going to make a video
[01:05:01] on making an AI agent um piano
[01:05:06] instructor.
[01:05:08] Add this to my Notion and make the intro
[01:05:12] for it. Uh make the intro for it based
[01:05:16] on my
[01:05:18] uh long-form videos and based on what
[01:05:23] you know about me,
[01:05:25] know about me in memory. And what you
[01:05:28] can do here,
[01:05:30] um you can actually at mention this is
[01:05:32] kind of more advanced, but like you can
[01:05:33] just at mention a folder. And so it'll
[01:05:36] actually look in the memory folder. You
[01:05:37] don't have to do that. That's more if
[01:05:39] you like create markdown files cuz you
[01:05:41] can just add folders. Like this is just
[01:05:42] a full-on computer, right? I can go into
[01:05:45] uh Sorry, I'm going on in in some
[01:05:46] different directions, but I just want to
[01:05:47] make
[01:05:48] I just kind of want to show people that
[01:05:50] like this is like a full-on computer
[01:05:53] that you can just like add things to. So
[01:05:55] like let's say I these screenshots,
[01:05:57] let's say these are important. You just
[01:05:58] drag them in.
[01:05:59] And so now you have a computer and you
[01:06:01] can tell the agent to analyze all of
[01:06:03] your photos and and um rename them based
[01:06:05] on what they are. You know, you can tell
[01:06:07] the agent to do that. Um but yeah, this
[01:06:10] is just kind of a segue. It's like this
[01:06:11] is just a full-on computer. Um
[01:06:15] but yeah, now it's just going to control
[01:06:17] notion. And so we can actually probably
[01:06:18] see it happen live. I actually don't
[01:06:21] know where it's going to add it.
[01:06:24] Um
[01:06:26] but the point is is it can control
[01:06:28] your notion. Okay, let me add this to
[01:06:30] notion and write the intro.
[01:06:33] Let's see where it adds it.
[01:06:35] >> You know, as it's doing this, the
[01:06:36] question that comes to mind, can you
[01:06:37] actually have these different agents
[01:06:40] interact with one another? So as an
[01:06:42] example, you have your comment analyzer
[01:06:46] on YouTube. Can that agent then almost
[01:06:50] like correspond with I don't know, my
[01:06:52] content planner for Instagram. Maybe I
[01:06:55] want to take a comment that I got on
[01:06:56] YouTube and use that as the basis to
[01:06:59] create a video that I'm going to post as
[01:07:01] a reel on Instagram. Like is there that
[01:07:03] level of integration yet between these
[01:07:06] different agents.
[01:07:09] >> Yes, and there's many ways to do this.
[01:07:11] First of all, I'll I'll get off this,
[01:07:13] but like you can see here, it created
[01:07:14] this Notion doc, right? Look at this.
[01:07:16] I said, AI P piano instructor. It took
[01:07:19] all of the things that it knows about
[01:07:21] me, about my content. And this is the
[01:07:23] exact, you know, obviously this is kind
[01:07:24] of a joke video idea. I'm not going to
[01:07:26] make this AI piano instructor, but it's
[01:07:28] like most people are using AI to write
[01:07:30] emails and summarize documents.
[01:07:31] Meanwhile, I just built an AI agent that
[01:07:33] can actually teach you how to play
[01:07:35] piano. It's like the exact vo- like it
[01:07:38] is my voice. It knows my how to make
[01:07:40] intros for me better than I do at this
[01:07:42] point. Um so, okay, that's kind of the
[01:07:45] first part, right? And it just gave me a
[01:07:46] link, view in Notion. And it allows me
[01:07:48] to just view it in Notion immediately,
[01:07:49] which is pretty cool. Um
[01:07:51] to answer your question. Your question
[01:07:53] was how to get these two communicate.
[01:07:56] In terms of how getting these agents, so
[01:07:58] like you just asked like
[01:08:00] if you want your YouTube analyzer agent
[01:08:03] and your Instagram
[01:08:05] >> Like content creator agent.
[01:08:07] >> content scripting agent.
[01:08:10] >> Yes, I would make those all one agent,
[01:08:12] first of all. So, I would make them the
[01:08:14] same agent with different skills and
[01:08:16] like each of these would be different
[01:08:17] tasks and different cron jobs, which
[01:08:19] we'll get to in just a second.
[01:08:21] Um
[01:08:22] So, if you make them the same agent,
[01:08:24] it'll just share the same memory, which
[01:08:26] is really cool. Like like it'll add
[01:08:28] things to the same memory, it'll update
[01:08:30] its soul.md file
[01:08:32] and it'll say, these are the list of
[01:08:33] things that I commonly do for Riley. Um
[01:08:36] and it will just go off and do them,
[01:08:37] which is really cool. Um
[01:08:39] and so, that's one way to do it. But if
[01:08:41] you have literally have separate agents,
[01:08:43] the thing that you can do is you can
[01:08:45] have them share a notebook. And so, in
[01:08:48] Notion, like you could make your agent
[01:08:50] log everything in Notion. Like so like
[01:08:53] let's say whatever your findings are,
[01:08:54] right? Let's say it generates a report,
[01:08:56] you can say, please upload this to
[01:08:57] Notion and put it in this database. You
[01:09:00] could have another agent that also
[01:09:02] stores things in the Notion database and
[01:09:04] it'll mark which agent is which, right?
[01:09:07] It'd say make sure to say to put CJ in
[01:09:10] for all of your entries. And so if you
[01:09:11] do that and you give your agents a
[01:09:13] shared notebook like on Notion, they
[01:09:16] both can use it and then they both learn
[01:09:18] from each other. And so that's how I'm
[01:09:20] seeing companies. I've talked to a lot
[01:09:21] of companies and trying to figure out
[01:09:22] how they do it. They have agents,
[01:09:24] different agents that share the same
[01:09:27] notebook. And so they're instructed to
[01:09:29] always check that
[01:09:31] every day and it will add relevant
[01:09:33] information to their computer as memory.
[01:09:35] Does that make sense?
[01:09:36] >> Yeah. And you know, to to build on that
[01:09:39] point because you said the way that you
[01:09:40] would do it, you would just all have it
[01:09:42] within one agent so that they have the
[01:09:44] same memory. And so do you you think
[01:09:47] about them almost these respective
[01:09:49] agents as like verticals? Like a company
[01:09:53] would have their content agent. And that
[01:09:56] would be working on it could be working
[01:09:58] on Instagram, TikTok, LinkedIn, YouTube,
[01:10:01] but it's like the content vertical. Then
[01:10:04] you might have, I don't know, the
[01:10:06] customer support agent. Is is that I
[01:10:10] guess in your companies and even for you
[01:10:11] personally, is that kind of how you
[01:10:14] separate the different agents in your
[01:10:15] mind?
[01:10:17] >> Yes. I I think I think
[01:10:20] I do separate them by vertical or or um
[01:10:23] I separate them by like
[01:10:26] what what context is useful to the
[01:10:28] agent, right? If
[01:10:30] if I have an Instagram agent, I think it
[01:10:33] would be very relevant to have shared
[01:10:35] context with anything about YouTube. And
[01:10:38] I would and or Twitter or anything,
[01:10:39] right? Because like you've probably
[01:10:41] figured this out that like what you do
[01:10:42] on YouTube, you can probably clip or
[01:10:45] take a graphic from and you can put it
[01:10:46] on Instagram. And so that those are that
[01:10:49] context is very relevant, right? I run a
[01:10:51] company
[01:10:52] and I have like I run a software company
[01:10:55] and I have a
[01:10:56] um
[01:10:57] and I run content. I use content to help
[01:11:01] the business, but sometimes something
[01:11:02] like customer support is not relevant to
[01:11:05] my Instagram, right? So, those agents
[01:11:07] would probably be separate, right? That
[01:11:09] context would only confuse the customer
[01:11:11] support agent. What's important for the
[01:11:13] customer support agent is it has all of
[01:11:15] the necessary information on how to help
[01:11:17] people when they submit a request on
[01:11:20] like how do I fix my app or you know,
[01:11:23] like how do I deploy my application?
[01:11:25] Right? You want to make sure that the
[01:11:26] information it has access to is relevant
[01:11:28] to that specific goal. And
[01:11:31] yeah, and like you want to think about
[01:11:32] you want to have
[01:11:34] if your agent has multiple goals, you
[01:11:36] want to make sure that they're aligned,
[01:11:37] right? Like if if you have a content
[01:11:39] creator agent, you might have five
[01:11:40] goals. Grow YouTube, grow LinkedIn, grow
[01:11:42] Twitter, you know, those kind of move in
[01:11:44] unison. If you're growing on Instagram,
[01:11:46] it's probably more useful to grow on
[01:11:47] TikTok. So, you want kind of these like
[01:11:48] aligned
[01:11:50] goals for your agent. And if they have
[01:11:52] completely separate goals where they're
[01:11:54] just like independent of each other,
[01:11:56] should probably be a separate agent, in
[01:11:57] my opinion.
[01:11:58] >> That's really valuable in terms of how
[01:12:00] to almost define and plan out your
[01:12:03] agents. It's like shared context, shared
[01:12:06] goals.
[01:12:07] >> Yes.
[01:12:09] Yes, exactly. You got it.
[01:12:11] And then yeah, it's just like making
[01:12:12] sure that like the things that your
[01:12:13] agent does are useful for that goal. And
[01:12:16] so, I think to kind of
[01:12:18] the one thing I do want to talk about
[01:12:19] before we go is I do want to talk about
[01:12:21] cron jobs. So, like here we have no cron
[01:12:23] jobs. Cron job you can think of as an
[01:12:25] automation trigger, right? This is just
[01:12:28] like any N or Zapier except you're
[01:12:30] triggering an agent to work. So,
[01:12:33] like we like that example earlier, I'm
[01:12:35] going to say
[01:12:36] okay, I want you okay, say okay, now
[01:12:41] that you have access to Supa data,
[01:12:45] Notion, and
[01:12:48] Supa data, Notion, and
[01:12:52] um my email.
[01:12:55] I want you to generate
[01:12:58] a report every morning. I'll zoom in
[01:13:02] here just so people can see a little bit
[01:13:03] better.
[01:13:05] Uh generate a report every morning that
[01:13:08] helps me
[01:13:10] come up with ideas for YouTube.
[01:13:14] Be creative.
[01:13:16] Generate a report.
[01:13:18] Um
[01:13:19] Look at YouTube comments
[01:13:23] if you need to. And that's part of super
[01:13:24] data or serp. There's another skill in
[01:13:27] here. If we go to marketplace,
[01:13:30] um here we have the serp API, um which
[01:13:34] will actually allow us to get YouTube
[01:13:37] comments. And again, these are all
[01:13:39] important things to know. It's like what
[01:13:41] tool and this is just an API. And we
[01:13:43] don't need to talk about what an API is,
[01:13:45] but like learning what an API is will
[01:13:47] allow you to get relevant context. So,
[01:13:49] all of this is relevant. Um
[01:13:51] if you need. Um please send me a report,
[01:13:56] a
[01:13:57] um markdown file, and public
[01:14:01] uh link. Um
[01:14:03] markdown file stored in
[01:14:07] documents. And right, we can uh mention
[01:14:08] documents.
[01:14:10] Um I like to just kind of um
[01:14:15] uh and public link
[01:14:17] uh of this report.
[01:14:19] Please create
[01:14:21] >> And to be clear, most of the context
[01:14:23] it's using to generate this report is
[01:14:26] just the fact that it knows your
[01:14:28] channel. Like it just has the link to
[01:14:30] your channel and I guess it has the
[01:14:32] access to your Notion.
[01:14:34] >> Yeah, yeah, yeah. And and that's likely
[01:14:36] all it needs, right? If it had access to
[01:14:38] my Notion, it would know who I am and it
[01:14:39] would be able to find that information.
[01:14:41] And you'd be surprised like what it's
[01:14:43] able to like connect the dots. Like
[01:14:45] it'll fill in the blank. So, like all I
[01:14:47] It probably would have got it just with
[01:14:49] me connecting my email. Like it would
[01:14:50] have been like, "Okay, his name is Riley
[01:14:51] Brown. Like I'll look him up on YouTube.
[01:14:53] Okay, this is his YouTube channel. Okay,
[01:14:55] like and based on the YouTube channel,
[01:14:57] it'll just gather its interests." You
[01:14:59] know, if I were to say like, "Hey, I
[01:15:00] want you to write a script like me." It
[01:15:01] probably would have just figured out
[01:15:02] like, "Okay, it's Riley Brown. Here's
[01:15:04] his YouTube channel. Let's go check it
[01:15:06] out." And then it probably would have
[01:15:07] been if At this point, it probably would
[01:15:09] have come back to me and been like,
[01:15:10] "Hey, um
[01:15:12] it would be super useful if you gave me
[01:15:13] access to the YouTube transcripts. You
[01:15:15] should download the Super Data API uh
[01:15:17] skill." And it would just do that.
[01:15:19] >> I And I And in my mind, it's like the
[01:15:22] I'm so glad that you shared it in the
[01:15:24] beginning. It's the loop. That it's like
[01:15:26] using that loop to get to those
[01:15:29] insights. And it's interesting because
[01:15:32] what I thought coming into the
[01:15:33] conversation is is that you would need a
[01:15:36] weekend or weeks in order to set this up
[01:15:39] to the point that it could have valuable
[01:15:42] outputs and like outcomes for someone
[01:15:45] that's listening at home. But it sounds
[01:15:46] like I I don't know how how long would
[01:15:49] we even estimate
[01:15:50] it would take for someone to get set up
[01:15:54] within a vertical, like one workflow,
[01:15:58] and they could actually start generating
[01:16:00] like valuable, helpful outcomes.
[01:16:04] >> I mean,
[01:16:06] how many minutes? I mean, how long have
[01:16:07] we been doing this for? And I've been
[01:16:08] talking for most of it. Um And so, you
[01:16:11] you say useful in there. And remember
[01:16:13] what we talk what I what I talk about.
[01:16:14] It's like you should never expect the AI
[01:16:16] agent to just create something useful
[01:16:18] for you because what's useful for me
[01:16:20] might not be useful for you. Maybe
[01:16:21] you're just starting out on YouTube.
[01:16:23] Maybe you don't have 200,000 followers
[01:16:25] like me, and you might need it to do
[01:16:27] something slightly different. But the
[01:16:28] point is is like, "Okay, in this time,
[01:16:31] it we connected it to all the relevant
[01:16:32] information, and now we gave it a cron
[01:16:34] job, right? It can now do things." So,
[01:16:36] like every day, it is going like you can
[01:16:38] see the full prompt here, right? You're
[01:16:40] a CJ Riley Brown's YouTube growth
[01:16:41] strategist. But like you don't even need
[01:16:43] to look at this.
[01:16:44] Um
[01:16:46] Let me zoom out here for a sec. Um
[01:16:50] Yeah, we we created this and this will
[01:16:52] run every single day and what we can do
[01:16:56] is we can test it. And I'm just going to
[01:16:58] test it right now. We can hit run.
[01:17:00] All of your cron jobs will start a new
[01:17:03] thread and it will be thread cron. So
[01:17:06] here we can um it's it's initiating and
[01:17:10] it'll get started in just a second. But
[01:17:12] basically
[01:17:13] it will start doing this task and you
[01:17:16] can play it however many times you want.
[01:17:18] And so the goal is
[01:17:20] if it's not super useful immediately,
[01:17:22] what you do is you can run it. It'll
[01:17:23] generate the full report, you can look
[01:17:25] at it and be like, "Okay, I you know,
[01:17:27] please edit the cron job or the it just
[01:17:30] you can say edit the cron job to make it
[01:17:32] so that the report is more like this."
[01:17:34] And then the agent be like, "Okay, I I I
[01:17:35] just changed the cron job. It's now more
[01:17:37] like this."
[01:17:38] And then you can go back to run, you can
[01:17:40] run it again, you can look at it, you
[01:17:42] can be like, "Yes, this is useful. Keep
[01:17:44] doing this every day." And what I like
[01:17:46] to do is I just like to set up a bunch
[01:17:48] of like automations like this where the
[01:17:49] agent will go off and do things. And
[01:17:51] every single time I I read one, I say
[01:17:54] like, "Hey, from now on I want it to be
[01:17:56] like this." Just like a new an employee
[01:17:58] that's receptive and good at kind of um
[01:18:01] adapting, the next time it creates this
[01:18:03] report, it will do that thing. And so
[01:18:05] that's
[01:18:06] the name of the game here in my opinion.
[01:18:08] It's just like
[01:18:09] giving it
[01:18:10] do having it do something and say, "Yes,
[01:18:12] that's good." or "No, you can do better.
[01:18:14] Do it like this."
[01:18:15] >> And so when you when you say do it like
[01:18:17] this, an example would be I want a
[01:18:20] summary at the I want this report in the
[01:18:22] format of like a summary on the at the
[01:18:24] top and then bullet points underneath
[01:18:26] and I want it to only be on one page
[01:18:28] maximum. Like that level of feedback. Is
[01:18:31] is that what you're you mean?
[01:18:32] >> Yeah.
[01:18:33] 100% and another thing that you can do
[01:18:35] if you've ever hired someone to make a
[01:18:37] report like this,
[01:18:38] um
[01:18:39] you can just upload the PDF straight to
[01:18:42] the computer and then and then tell the
[01:18:43] or you can upload it straight to the
[01:18:45] chat and be like, I want it to look like
[01:18:47] this. And what it will do is, in the
[01:18:49] skills folder, right? In the If we go to
[01:18:51] the files folder and we go into skills,
[01:18:54] if you like click into, let's say,
[01:18:56] YouTube competitor analysis,
[01:18:59] here, this is just a skill.md file. But
[01:19:02] what you can also do in the same
[01:19:04] in this folder that has the skill.md
[01:19:07] file, you can add a folder and that you
[01:19:09] could put like references. And then in
[01:19:11] the skill, you could say, "Please look
[01:19:13] at the references whenever you use this
[01:19:14] skill." And it will just generate it
[01:19:16] exactly like whatever is in the
[01:19:18] references. You know what I mean? Like
[01:19:19] you can give it an example and put it in
[01:19:21] the skill so that it always looks a
[01:19:23] certain way. Um which which I think is
[01:19:26] really really useful. But you can see
[01:19:27] here that the the the cron job
[01:19:29] automatically started. This one is going
[01:19:32] to take a while, right? It's still
[01:19:33] going, but as soon as it's done, I can
[01:19:36] evaluate it and then tell the agent how
[01:19:37] to improve it in the future.
[01:19:39] >> Interesting. So you're you're creating
[01:19:42] the automation, you're then having it
[01:19:45] fulfill the task, and then once you get
[01:19:48] the output, you're giving it feedback on
[01:19:50] the cron job so that it can improve it
[01:19:52] for the next iteration.
[01:19:55] >> Yes. And here you can see,
[01:19:56] Open-interpreter just dropped a big
[01:19:57] release
[01:19:59] with task flows. I don't even know what
[01:20:01] this is. This is super useful. Like
[01:20:04] uh apparently Open-interpreter had a new
[01:20:05] release and it's grabbing more details.
[01:20:08] It'll get back to me when it's done.
[01:20:10] >> Yeah.
[01:20:11] You know, before we get out of here,
[01:20:12] with these cron jobs,
[01:20:15] how wacky and almost customized
[01:20:19] can I make it? If I wanted to instruct
[01:20:21] it that like after you create this
[01:20:23] report, I want you to send me an email
[01:20:26] at 9:00 a.m. every morning with the
[01:20:29] report like attached in that email so I
[01:20:32] can review it when I wake up. Like can
[01:20:34] you get to that level of granularity?
[01:20:38] >> You can go 10 times
[01:20:41] above that. You know, you know, going
[01:20:43] back to this right here, like you
[01:20:45] basically want it to like generate the
[01:20:48] report and then send you an email. It
[01:20:50] can absolutely do that. And then it
[01:20:51] could generate a report, send an email,
[01:20:52] and then it can send an email to
[01:20:54] everyone at your company.
[01:20:55] Um
[01:20:56] you can absolutely do that.
[01:20:57] >> Yeah. You know what, even as you were
[01:20:59] talking, kind of where I've landed with
[01:21:01] all of this Riley, and and I'm not as
[01:21:03] experienced as you,
[01:21:05] is
[01:21:06] we almost we just have to start thinking
[01:21:08] about these AI agents and its
[01:21:09] capabilities as like an employee. And
[01:21:13] anything it feels like anything that an
[01:21:15] employee could do,
[01:21:17] it's getting to the point where the AI
[01:21:19] has the capability to do that in a
[01:21:22] pretty autonomous
[01:21:24] fashion.
[01:21:26] If that's the case, I'm just curious cuz
[01:21:28] I know that you're an optimistic
[01:21:30] >> guy.
[01:21:31] >> And we hear so much of like the dooms
[01:21:32] day scenarios with AI.
[01:21:35] What do you think that that means
[01:21:38] for people that today in 2026, we're
[01:21:42] already at a point, and it's early like
[01:21:45] you mentioned,
[01:21:46] where
[01:21:47] it feels like open claw can handle much
[01:21:50] of the capabilities of just like an
[01:21:53] employee, but it just runs itself and it
[01:21:56] can run for longer and longer amounts of
[01:21:58] time.
[01:22:02] >> What are the What are the implications
[01:22:03] of it of it running for a really long
[01:22:05] period of time? Um
[01:22:07] I would say
[01:22:09] I would say that of course there you
[01:22:12] know, I I I I I I I I I I I I I I I I I
[01:22:13] I I I I I I I I I I I I I I I I I I I I
[01:22:13] I I I
[01:22:15] majority optimist. It's important to
[01:22:17] never like put myself into a bucket
[01:22:19] either optimist or pessimist. I think
[01:22:21] it's good to kind of be like 70/30,
[01:22:23] right? You want to be a realist, right?
[01:22:25] You don't want to just kind of like
[01:22:27] you know, close your eyes to bad news
[01:22:29] and and uh bad
[01:22:31] uh like when things are bad. And I will
[01:22:32] say in the short term there will be many
[01:22:34] people losing their jobs to this.
[01:22:37] I think the people who learn this will
[01:22:38] be better to weather the storm. I think
[01:22:41] if many CEOs will be pressured to
[01:22:45] replace a lot of their work and save
[01:22:47] money, right? When when new technology
[01:22:49] comes out and CEOs can save money, they
[01:22:52] will do it, right? Because they're or
[01:22:53] else their competitors will do it and
[01:22:55] then they'll put them out of business,
[01:22:56] right? It's kind of they're pressured in
[01:22:58] doing it. And so if you can prove to the
[01:23:00] company you work at that you are able to
[01:23:03] think clearly and create agentic
[01:23:05] workflows for the company that can help
[01:23:06] the company, um it's usually the I think
[01:23:10] you'll be way better off.
[01:23:12] Um and I also think that many people,
[01:23:16] many really smart people who might be
[01:23:18] making 500k at Meta, you know, they
[01:23:20] might be making a ton of money at
[01:23:22] Google, will be laid off. And I think a
[01:23:24] lot of these people will start
[01:23:25] companies. And I think it will be a lot
[01:23:27] easier to actually start companies and
[01:23:30] uh as a team of two or three and get the
[01:23:32] work done of of a team of 20 in the past
[01:23:35] because these and if you have like a
[01:23:37] little bit of startup capital and you're
[01:23:39] able to front the the costs for the um
[01:23:43] the tokens, I think you'll be able to do
[01:23:45] a ton with a few people. So I think
[01:23:46] we're going to see a bunch of
[01:23:50] a rise in smaller like startup companies
[01:23:53] that are like a handful of people. And I
[01:23:55] think this is really fun and I think
[01:23:57] those people like you want to be very
[01:23:58] high in openness. You want to be
[01:24:01] like it's it's helpful to be smarter and
[01:24:03] it's helpful to just be if you're really
[01:24:05] good at managing a lot of people at the
[01:24:07] same time, you're probably going to be
[01:24:08] good at managing a lot of AIs at the
[01:24:10] same time. So I think getting really
[01:24:12] good at stating outcomes that you want,
[01:24:15] um delegating to AI agents and like you
[01:24:17] said, going through these evaluation
[01:24:19] processes and making sure that it's not
[01:24:21] just AI slop. Um and they're actual
[01:24:23] things that they can help. Um if you
[01:24:25] want to create an agent that handles
[01:24:26] email, like you should probably test
[01:24:28] that a lot, especially if there's
[01:24:30] clients on the other side or potential
[01:24:32] clients that you can close the deal. You
[01:24:34] want to make sure that these workflows
[01:24:36] are set up in a way where it's really
[01:24:37] high quality. And so, I would really
[01:24:39] focus on that. Building agents that are
[01:24:41] really high quality, getting in a loop,
[01:24:43] improving them, making them better. And
[01:24:46] to do that, you need to really
[01:24:47] understand a certain niche or industry.
[01:24:50] You know, I really understand YouTube
[01:24:52] and content creation. So, I'm able to
[01:24:53] create really good agents. Someone who
[01:24:55] has no idea how to create a hook for Tik
[01:24:58] Tok or YouTube, they're not going to be
[01:25:00] as good at creating an agent for content
[01:25:02] creation. So, you kind of want to learn
[01:25:04] a a little bit about a lot of
[01:25:06] um uh
[01:25:07] any area that's useful
[01:25:09] or any area that you're trying to create
[01:25:11] agents for, you do kind of want to be a
[01:25:12] domain expert or else you don't know
[01:25:14] what AI slop is. You you know what I
[01:25:16] mean?
[01:25:17] >> Yeah, I'm I'm so glad that you made that
[01:25:19] point. Which is like
[01:25:21] your your career isn't wasted. Like the
[01:25:24] fact that you have this level of
[01:25:25] expertise. Cuz I think sometimes that's
[01:25:28] how it's presented with this AI stuff.
[01:25:29] It's like, "Oh, it just replaces you and
[01:25:31] it's over." It's like, "No, your
[01:25:33] expertise is still valuable. You're just
[01:25:36] using it in collaboration with the AI."
[01:25:38] And
[01:25:39] one one thing that I think um
[01:25:42] I'm such a big believer in, Riley, is
[01:25:46] just getting to action. And like that
[01:25:48] first action step as quickly as
[01:25:50] possible.
[01:25:51] Because for the person listening at home
[01:25:53] and also for myself, it's like you start
[01:25:55] building momentum when you take that
[01:25:58] first action step. And so,
[01:26:01] Riley, if you almost had to give someone
[01:26:05] like now that you've watched this you've
[01:26:07] seen this conversation and you have all
[01:26:09] of these ideas and use cases,
[01:26:12] immediately after watching this video,
[01:26:15] what is the thing that you would advise
[01:26:18] or suggest that someone listening at
[01:26:21] home goes and does?
[01:26:24] >> Good question. So, what can someone What
[01:26:26] should someone do first, right?
[01:26:28] Uh if you're getting started with AI
[01:26:30] agents, the first thing that you should
[01:26:32] do is automate one single task, right?
[01:26:34] Pick one thing that's annoying to you
[01:26:38] and solve it with AI. Make it faster or
[01:26:40] better. You know, add a research step or
[01:26:42] add an automation step. There's a lot of
[01:26:44] things that you know, like the one thing
[01:26:47] that I did that gave me immediate
[01:26:49] uh value is I automated having a social
[01:26:52] media manager, right? When a company
[01:26:55] reaches out to me, they offer me money
[01:26:57] for a brand deal or or etc., the first
[01:26:59] thing that the agent does is the agent
[01:27:00] will analyze who that company is, and it
[01:27:03] will take a look and see like is it a
[01:27:04] legitimate company or is it not? It'll
[01:27:06] analyze the video of the companies that
[01:27:07] I've worked with in the past to see if
[01:27:09] it's the type of company I want to work
[01:27:11] with, and then if it is, it'll just
[01:27:12] respond with and say like hey, like what
[01:27:14] are your rates? I try to get them to
[01:27:16] share their budget, and it knows exactly
[01:27:18] how much I charge. And so, basically, it
[01:27:21] I've created a uh three separate cron
[01:27:25] jobs. Every day, the agent will go
[01:27:27] through and analyze it and send all of
[01:27:28] the emails that it needs to
[01:27:31] um based on the entire inbox, and it
[01:27:33] will archive a lot of the ones that like
[01:27:35] aren't that good. And then it sends me a
[01:27:37] little brief that says like um okay,
[01:27:40] these companies don't seem to be ones
[01:27:42] that you'd be interested in, and then
[01:27:44] and it'll say like here are the ones
[01:27:45] that I've talked to. And then if it's
[01:27:47] not super confident, it'll write a draft
[01:27:49] for me and say like hey, like do you
[01:27:51] want to send If you want to send this,
[01:27:52] here's a link to the draft in um
[01:27:55] in
[01:27:56] Gmail, and it will send me a link to it,
[01:27:58] and then I'll go in and send the draft.
[01:28:00] And so, that's one example of something
[01:28:02] that's really, really useful. That one
[01:28:05] took me 3 days, three full probably 10
[01:28:07] plus hours to set up because
[01:28:11] as I mean, think about it. I mean,
[01:28:12] that's that's a pretty hard task to get
[01:28:14] anyone to do and there's a ton of rules.
[01:28:17] So, I mean, I had to make a fake email
[01:28:19] to send in fake requests uh to get to
[01:28:23] test to test it and sometimes you do you
[01:28:24] have to do that. You have to like create
[01:28:26] a little testing environment where you
[01:28:27] can actually test the agent and I had to
[01:28:29] set up a separate email for it so I
[01:28:31] could send it fake emails and then I
[01:28:33] would go to the cron job and I would
[01:28:35] just hit run to see how it would do it
[01:28:37] and then after
[01:28:38] many hours, I got to a point where I'm
[01:28:40] like, yeah, 80 95% of the time this
[01:28:42] thing says the right thing.
[01:28:43] >> You You said 80-90% of the time and I
[01:28:46] actually think cuz what you just shared
[01:28:47] is is so valuable. But the I think about
[01:28:50] the unintended consequences.
[01:28:53] Like the I've I've heard use cases of
[01:28:55] oh, it deleted this sheet or this piece
[01:28:58] of data that I needed or it sent this
[01:29:01] email. I didn't want it to send this
[01:29:02] email. Can you share like your
[01:29:05] experience of unintended consequences
[01:29:07] using these agents because they're
[01:29:10] autonomous?
[01:29:11] And then also how you've like reduced
[01:29:13] that
[01:29:14] instant those instances over time.
[01:29:17] >> Yes. Um so, first thing, um
[01:29:20] like I said, create a safe environment
[01:29:22] to do it. Like you would I like when I
[01:29:23] was first testing Notion, I created like
[01:29:25] a new Notion database with like fake
[01:29:28] data. Um what I did is you can actually
[01:29:30] ask the agent to duplicate a Notion So,
[01:29:32] if you wanted
[01:29:33] cuz
[01:29:34] the reason I started doing this is one
[01:29:36] time when I was using Notion I my agent
[01:29:38] deleted the file that I was working on
[01:29:40] and I couldn't figure out how to bring
[01:29:41] it back. And actually like it didn't
[01:29:43] update. I would have it would have been
[01:29:44] super annoying to bring it back so I had
[01:29:45] to start over and that was really
[01:29:47] frustrating. And so, and that was a
[01:29:49] human error. I forget what my error was.
[01:29:51] But then now whenever I'm setting up an
[01:29:53] automation, what I'll do is I'll just
[01:29:54] say, "Hey, can you duplicate this
[01:29:55] database and put it in a different team
[01:29:57] space?" And then I'll just test the
[01:29:59] agent there.
[01:30:01] Um and so, I would advise you to try
[01:30:03] wherever you can create like a dummy
[01:30:05] email or a dummy database on notion or a
[01:30:07] dummy Google um Drive folder to test to
[01:30:11] see how these workflows work so that if
[01:30:13] you if it does make a mistake, it
[01:30:15] doesn't matter, right? It's all fake.
[01:30:18] Like it's it's just dummy data or maybe
[01:30:19] it's just a duplicate of a folder.
[01:30:22] And then when things go wrong, which
[01:30:24] things will go wrong every once in a
[01:30:26] while. But like things go wrong with
[01:30:28] human employees as well, you know, like
[01:30:30] they make a mistake and it costs company
[01:30:31] time. It's just part of doing business
[01:30:34] with humans or agents. And so
[01:30:37] I think
[01:30:39] I think the best thing to do is just to
[01:30:40] really get clear about what you want.
[01:30:42] And you can even specify in your
[01:30:44] instructions in your skills or if you're
[01:30:46] typing a prompt, like literally just
[01:30:49] um
[01:30:50] just say like don't do these things. And
[01:30:53] if you catch the mistakes that it does,
[01:30:56] what you can do you'll you'll start to
[01:30:58] learn what to tell it not to do, which
[01:31:00] is really important. And so
[01:31:02] to kind of
[01:31:04] like to conclude here, like the more you
[01:31:06] work with agents,
[01:31:08] uh the better you get at working with
[01:31:10] them. There's no list of rules. Part of
[01:31:12] this is just kind of interacting with
[01:31:14] agents, making sure they get better, and
[01:31:16] then over time you're just going to get
[01:31:17] better at
[01:31:18] um
[01:31:19] over time you're just going to get
[01:31:20] better at telling them what not to do,
[01:31:21] what to do, how to give them
[01:31:23] instructions. So it's something that you
[01:31:24] learn as you do it, not necessarily just
[01:31:26] like following a rule book, which is why
[01:31:28] it's the most fun skill to learn right
[01:31:29] now in my opinion.
[01:31:30] >> Yeah. By the sounds of it, Riley, and
[01:31:32] thank you so much for all the
[01:31:34] information you've shared, it's like it
[01:31:36] feels like a conversation like this gets
[01:31:39] you started, and it shows you what's
[01:31:41] possible, and some of the use cases, and
[01:31:44] how to get set up. But then it's really
[01:31:47] a case of you just have to spend time
[01:31:50] actually building, and trying, and
[01:31:53] iterating, and it fails, and then you
[01:31:55] iterate again, and then maybe it makes a
[01:31:57] mistake and like deletes something, and
[01:31:59] you correct off that. That feels like
[01:32:02] the the process of how you get to a
[01:32:04] place where this is actually valuable.
[01:32:08] >> 100% Yes, I I I couldn't agree more.
[01:32:11] Yeah, it
[01:32:13] um
[01:32:13] yeah, 100%.
[01:32:15] >> Yeah, before we get out of here, Riley,
[01:32:17] is there anything I think about the last
[01:32:19] time that you and I recorded hundreds of
[01:32:21] thousands of people, over 200,000 people
[01:32:24] found out about Vibe Coding and what was
[01:32:26] possible there for the first time.
[01:32:29] Knowing obviously we don't know where
[01:32:31] this conversation will go, but is there
[01:32:33] anything that you want to say before we
[01:32:34] get out of here?
[01:32:36] >> You know, one thing I did remember about
[01:32:37] our last conversation, I promised
[01:32:39] everyone, I said, "Hey, if you make an
[01:32:41] app on any software, doesn't matter
[01:32:43] which one, and you shared it on Twitter,
[01:32:45] I'll give you a retweet."
[01:32:47] Um
[01:32:47] and just tag me or tag you in it. Um
[01:32:50] I think I'll let's do the same thing. I
[01:32:52] think
[01:32:53] I want to challenge you guys to do one
[01:32:55] specific agent workflow, and you can
[01:32:58] either screenshot like a cron job, or
[01:33:00] you can screenshot the input, you can
[01:33:02] just put the input and the output and
[01:33:03] say like I created an agent that solves
[01:33:05] this problem, and it saves me x amount
[01:33:08] of money or x amount of dollars. I think
[01:33:10] it would be really useful and fun if you
[01:33:12] guys posted that on X or something.
[01:33:14] That's where I learn the most about AI.
[01:33:16] If you guys are on X, um just tag me in
[01:33:19] that and I'll I'll give it a retweet if
[01:33:20] it's a good
[01:33:21] uh workflow. I think that would be fun.
[01:33:23] We did that last time.
[01:33:24] >> That's awesome. Riley, you're a star,
[01:33:26] mate. Thank you so much.
[01:33:28] >> Of course. This was fun. Appreciate it.
[01:33:30] >> This was awesome. So, if you enjoyed
[01:33:32] this conversation and you want to hear
[01:33:34] even more stories like this, then just
[01:33:37] click here. And also, my team is going
[01:33:39] to put some more videos that you can
[01:33:41] watch here. Thank you.