Why Specialized Agents are Superior (How I Built an OpenClaw Superteam)

0:00 So, I spent the last 2 weeks building hundreds of different AI agent workflows,

0:03 mostly using Open Claw.

0:05 But, I also use Manas Claw code and even Perplexity computer,

0:08 which just came out.

0:09 And my biggest realization through this process,

0:12 companies are going to have very narrow AI agents that operate in a team.

0:17 And my current plan is to build 15 high-quality AI

0:21 agents that run our entire growth division here at vibecheck.dev.

0:27 And so, I want to take some time in this video

0:30 to explain why I believe that narrow agents are the future.

0:34 And we'll also kind of talk about why I'll be using Open Claw for this project.

0:39 And so, let's just dive into the video.

0:41 Over the past 2 weeks, we tested many different agents.

0:44 And the main four that we tested were Open Claw,

0:48 we tested Manas, we tested Claw code, and we tested Perplexity computer.

0:54 And so, Perplexity computer is actually really interesting.

0:57 You see here, you can actually switch from search to computer.

1:01 And the way Perplexity computer works, I can say,

1:04 "Please make an app." This right here is a single task I can give an AI agent,

1:12 and just like ChatGPT, it will start to work.

1:14 Except, this AI agent will get a sandbox,

1:18 which is just a computer that's running in the cloud,

1:21 and it can actually create files.

1:23 You can see here that whatever it creates

1:26 can open up in this side panel right here.

1:29 So, it's like ChatGPT with a computer, which is exactly like Manas.

1:34 And Manas has been around a lot longer than Perplexity computer,

1:37 and it operates the same way.

1:39 Manas was the first kind of general agent tool that was released

1:43 that had every single task that you put in has access to a computer.

1:48 And as you can see here, you can see view Manas's computer.

1:51 And so, we can view this over here.

1:53 And so, you can see that this AI agent comes with a computer,

1:57 it can create files, it can edit files, and it can do many

2:00 different things that you would do on a computer.

2:03 And so, that's how Manas and Perplexity computer work.

2:06 You enter a task, it spins up a computer,

2:09 and depending on how you prompt that task,

2:11 different things will happen on that computer.

2:13 It can create different things, it can do a whole host of things.

2:16 And so, if you run five tasks, each one comes with its own little computer.

2:21 And so, this can be seen of more as like a command center.

2:25 Right?

2:25 This is a command center for agents that have access to a computer.

2:33 And this is cool for certain things, but it's actually not what we want.

2:38 I don't believe this is going to be the most useful form of AI agents.

2:43 I think the most useful type of AI

2:45 agent will be something exactly like Openclaw.

2:49 Openclaw is an AI agent that runs on one computer.

2:53 And you can see that uh Mac Minis are literally sold out right now.

2:57 It's really hard to get a Mac Mini or a Mac

2:59 Studio because so many people are running an AI agent,

3:03 Openclaw, on these computers.

3:06 And so, basically, what Openclaw did is they

3:08 basically put an AI agent on a computer,

3:11 and then they gave it really good memory,

3:14 they gave it really structured skills that you could very easily add,

3:19 and then they also added a gateway.

3:21 And this gateway allowed you to chat with Openclaw from different applications.

3:27 You could do it on Telegram, you could do it on WhatsApp,

3:31 you could message it on Discord, on Slack.

3:35 And this is why Openclaw went really viral.

3:38 It gave an AI agent a computer, and then made it accessible in all

3:43 of the tools that you already communicate with other people.

3:46 And so, the first Openclaw AI agent that I created had many skills.

3:52 So, I and so this is an overview of the most

3:56 useful skills that I gave my first AI agent.

3:59 My favorite skill was this social media

4:01 transcript analyzer using an API called Super data.

4:05 If you guys want to look it up and use it,

4:07 it allows you to turn any YouTube link, Twitter link, Instagram link,

4:12 or TikTok link into a transcript so you could very easily analyze social media.

4:19 I added this.

4:20 And then I added the ability for it to control my notion.

4:23 And then I added the ability to control all of my Google workspace.

4:29 So, my calendar, my email, Google Docs, Google Sheets, etc.

4:33 And then I gave it access to our linear so I could

4:36 take a look and see where are we at with the product,

4:39 what's launching soon, things like that.

4:42 I gave it access to Figma.

4:43 It could literally control Figma on my computer.

4:46 It could generate any type of media using FAL.

4:49 It could even edit videos, which was my previous video.

4:52 And what I realized over time is that the more skills that I added, right?

4:56 As the amount of skills increased, the dependability of the AI agent decreased.

5:05 And so that's when I realized, okay,

5:06 well, you can't really add unlimited skills.

5:08 It stops being super useful.

5:10 It doesn't use the skills at the right time.

5:12 The context gets super clouded and it doesn't use

5:16 the right integrations and the personalities ended up getting jumbled.

5:20 And so that's the conclusion that I came to, right?

5:23 We need to create a team of AI agents that have,

5:28 I would say, seven to 10 skills each.

5:32 Instead of building out AI agents that have 30 skills.

5:36 This is the sweet spot.

5:37 As you go above this, the AI agent stops performing super well.

5:42 And so that's one of the main reasons why I think Manis has a lot of potential.

5:45 It's just not my tool of choice because when you hit use skills,

5:49 you can go to the manage skills and you

5:51 can kind of see all of their official skills.

5:53 And so, they have all of these different types of skills that you can add,

5:58 but you're not adding it to a specific agent.

6:00 Rather, you're adding it you're adding it to your command center,

6:04 which means everything is proactive.

6:06 You have to go to your command center and then ask it to do it.

6:09 People simply want an employee that gets things done.

6:15 And if you think about a really good employee,

6:17 you think like the employee will actually just like do things that surprise you.

6:22 A good employee will make suggestions that are useful.

6:27 And I believe that in order to do these three things,

6:30 like get things done, do things that surprise you,

6:32 and make suggestions that are useful, you need to have specific goals,

6:39 or you need to give AI agents intent.

6:42 And I actually got this from on Twitter, Emmett Shear.

6:45 He was the interim CEO of OpenAI when Sam Altman almost got fired,

6:49 but he tweeted "Prompts are so late 2025.

6:53 We are giving models intents now." And I think I I

6:57 would say like we are giving AI agents with computers intents.

7:02 And the definition of intent is intention or purpose.

7:06 Right?

7:07 We are giving these agents purpose.

7:09 And I believe that if you're going to go through this paradigm,

7:12 it's really hard to give these agents purpose because they're so general.

7:16 They have so many different skills.

7:18 It's super proactive.

7:20 Uh and so, that's why I don't really like Perplexity computer and Manas.

7:24 And so, that's why I want to create a team

7:26 of narrow open claw agents with very specific goals and skills.

7:32 And I'll explain a little bit more uh about why I want to do this.

7:35 And so, when we were testing these AI agents,

7:38 and I have a bunch of these AI agents running, this is my journal bot.

7:42 And uh I have these agents running in Telegram right now.

7:46 And after testing them for 2 weeks,

7:48 I realized that this is the way this whole space is going.

7:51 A focused agent with a specific personality,

7:54 with specific tasks, and a specific heartbeat.

7:59 And what I realized is that all

8:01 of these things are better when they're focused, right?

8:03 When they're when you have a team of AI agents

8:05 or a team of agents running on a computer that are confined right?

8:10 To a more narrow focus, everything performs better,

8:14 which allows you to focus your skills and integrations

8:18 on what will be useful to reach a specific goal.

8:21 Let me give you an example.

8:22 So, my favorite agent that I'm using right now

8:25 that I message in Telegram is this content bot,

8:29 which is specific for creating YouTube videos.

8:33 This is my YouTube agent.

8:35 So, this focused agent is my YouTube AI agent.

8:40 And the only thing that this agent is focused on is creating YouTube videos.

8:45 And it has three goals in its files.

8:49 It knows exactly what goals it's optimizing for, which are subs,

8:53 views, and conversions.

8:55 And I'm not to say that everything that I

8:57 do on YouTube will is optimized for these three things,

9:01 but whenever I ask it to create a script, for example,

9:04 it knows that I want to increase the amount of subs,

9:07 increase the views, and increase conversions.

9:10 So, the main reason really narrow goals are super useful is that it allows

9:14 you to create hyper-specific skills that can

9:18 be verified whether you should add them.

9:20 If your skill does not have anything to do

9:23 with your with your goals of your AI agents, you shouldn't add them.

9:27 So, it makes it super easy to add skills.

9:29 And for example, the main skill that I

9:31 use for this YouTube AI agent is YouTube research.

9:34 And then this allows me to think about, "Okay,

9:36 what integrations are useful for this YouTube research

9:40 skill?" And so, I use two for this.

9:42 I use the SERP API skill integration, and I also use the Super Data API.

9:48 This one allows you to scrape transcripts.

9:50 This one allows you to like search through YouTube.

9:53 It's really useful.

9:54 Skill number two is thumbnail generator, right?

9:57 I generate thumbnails, and I even have my AI agent

10:01 every single morning scrape my competitors' thumbnails,

10:05 and then it comes up with ideas and kind

10:07 of modifications of their videos with my face.

10:10 And so, for this, we need access to Nano Banana.

10:13 And for this skill specifically, it it has some relevant context that it needs,

10:18 which is photos of Riley, right?

10:22 It needs my photos in order to create an image of me.

10:25 So, that's just some useful context that we need to give it.

10:28 And number three is it needs to be able to control my Notion,

10:34 which is where I keep all of my scripts for my YouTube videos.

10:39 And for this, we actually just need the Notion integration.

10:43 And so, you can see here that it's kind of this direct path, right?

10:46 Your YouTube agent is in charge of optimizing for YouTube subs.

10:51 It wants to increase the amount of views you get,

10:54 and it increases the amount of conversions you get from your video.

10:57 That is what my agent is optimizing for.

11:00 And now, when I go to my agent and say,

11:01 "What skills do you need?" it knows exactly where we're going, right?

11:06 This is a path.

11:07 And if you've spent any time hiring people,

11:11 the most annoying people to hire are people with vague skills.

11:14 They don't have specific goals.

11:15 They're good at talking,

11:17 but they're they're not good at driving towards a specific goal.

11:20 They're good at distracting from the goal.

11:21 The best employees to hire are people who are like,

11:24 "Yep, I'm really good at certain things,

11:27 and I can help your company reach these goals,

11:29 which will ultimately help your company." It's very simple.

11:33 And this YouTube agent is just one of the agents that I use.

11:37 And so, the question now becomes, why narrow agents?

11:42 The first reason why is when we find a super useful agent,

11:46 it's very easy to duplicate, right?

11:48 You can remix it for something else.

11:51 It could be relatively simple to turn

11:53 a YouTube agent into a TikTok agent, right?

11:55 We could create a TikTok agent that's only focused on TikTok.

11:58 We could create one that's only focused on Substack, for example.

12:02 And when you create these smaller agents, it's easier to duplicate.

12:06 Uh you know, when you try and create a massive agent that has like 50 skills,

12:11 it's just hard to extract just the the portion

12:15 uh that you want to duplicate out of it, right?

12:17 And so, I can very easily duplicate my my single um narrow-focus agents.

12:23 Additionally, um this makes it super easy to share with your team.

12:28 So, for example, today I built a journal agent.

12:32 My journal agent is it lives in Telegram,

12:35 and this agent is a little bit more hands-on.

12:38 And so, basically, what it does is it reaches out to me every 30 minutes.

12:42 Sometimes it doesn't reach out if nothing needs to be done,

12:45 and it analyzes everything that I do,

12:47 every meeting, every video that I make, everything.

12:50 And if it wants more context, it'll ask me.

12:52 Every single day it'll write multiple journal entries,

12:55 logging everything that's useful um and everything

12:59 that it needs to know about me.

13:00 Just so I can create this like running log of all

13:03 of the important information that's important to business and content.

13:06 And the purpose of this agent is that it informs all of the other agents, right?

13:11 So, this journal agent has access to Notion.

13:14 Every single agent that we have has access to Notion,

13:16 and all of these other agents are aware

13:19 of the journal that the journal agent creates.

13:23 So, my email newsletter every day is just going to read my journal

13:27 agent's journal and it's going to come up with ideas for email newsletters.

13:31 So, in my journal agent, it'll know when there's product updates and then

13:34 my email newsletter agent will be like

13:37 will just draft up an email newsletter that needs to go out to our email list,

13:41 which is 300,000 people and in this email newsletter agent,

13:46 it has very specific goals, right?

13:47 This newsletter agent has very specific goals,

13:50 which is optimize the amount of conversions from our email newsletter, right?

13:54 And and to maximize click-through rate, to open rate, things like that.

13:59 And it doesn't have to be clouded by any of the journal agent's goals.

14:03 It just has access to the journal that the journal agent creates.

14:07 And so anyway, this is a really useful agent

14:09 that I created and I want to help Ange, my co-founder, create more content.

14:14 And so now, because I made a really narrow journal agent,

14:17 I actually just shared with him my entire Open Claw agent

14:21 and he was able to duplicate it in about 5 minutes.

14:24 So, it was really shareable.

14:26 So, when you create a really narrow agent that's super useful,

14:29 it's very easy to duplicate it and share it with other people.

14:32 And it also makes it just more understandable, right?

14:35 This Open Claw agent runs on a computer.

14:39 In the computer, it has access to files, right?

14:41 They like this computer, every time you use Open Claw,

14:44 you're basically editing files, right?

14:46 All of these skills are just markdown files stored in the Open Claw folder.

14:52 If your focused AI agent only has

14:54 a few skills and and a handful of integrations,

14:57 it's a lot easier to understand when you send them to other people.

14:59 And then finally, right?

15:01 This one's pretty intuitive, right?

15:02 If you have a narrow set of of goals, right?

15:06 Goals, right?

15:07 You want to hit specific KPIs.

15:08 In the case of the newsletter, it would be like open rate, subscriptions, right?

15:13 You want people to subscribe and ultimately like click-through rate.

15:17 And like if you were just had like a newsletter business,

15:19 it would be like, you know, you could have revenue.

15:22 When you have very narrow goals, it's very reviewable.

15:26 You can look at that agent and be like, "Yep, you did a good job." Or, "No,

15:29 you did a bad job." You know exactly where it needs to change.

15:32 You It's It's It's pass fail.

15:34 When your AI agents are pass fail, it's a lot easier to just cut them, right?

15:38 The A lot of AI agents that you create over

15:40 the next few years are not going to be worthwhile.

15:42 And the more narrow they are, uh the easier it is to say,

15:45 "Yep, you did good." Or, "No, you did bad,

15:47 so get rid of it." And then the final two reasons

15:49 you want a narrow focus is you can create easier loops,

15:53 which allows them to be more autonomous.

15:56 I have multiple narrow agents that are very simple loops, right?

16:00 It only has a set of three tasks that it does every single day.

16:03 It knows what it's optimizing for, and it can just go in those simple

16:07 loops over and over and over again because tasks are just are called cron jobs,

16:13 which are triggered at specific times during the day.

16:15 And the more narrow your agent, the easier it is to get in a predictable loop,

16:19 and you can just let it run.

16:20 If you have a super mega agent that's super massive, it's harder to do this.

16:24 So, I think you understand my objective here.

16:27 I want to create very narrow AI agents with very specific goals,

16:32 and I want them to operate in a team, right?

16:34 This could be my team.

16:36 It's hard to say exactly which agents I'll be adding, right?

16:39 As I do more workflows, I'll notice where we need to create a new agent.

16:43 But the one thing that I think that Perplexity and Manas

16:46 got right is they're actually using a computer in the cloud.

16:50 They spin a computer per task, right?

16:53 When you type in a task, they spin up a computer,

16:56 and that agent can use the computer.

16:58 I actually don't think that's going to be the paradigm.

17:00 I think it's going to be Open Claw running in a computer in the cloud.

17:04 And so, it's up to us as a company to figure out,

17:07 "How do we efficiently run these in the cloud?"

17:09 And 2 years from now when each one of them has,

17:12 you know, 20 agents, that's 200 AI agents.

17:15 How do we efficiently run all of these AI in the cloud?

17:18 Also, how do we share these AI agents with other people on the team?

17:21 That's one thing that we really need to figure out.

17:23 And then, how do we get these AI agents to be able to communicate one another,

17:27 or at least share memory?

17:30 And in future videos, I'll be talking about how to do this.

17:33 How to get your AI agents to actually share memory,

17:36 so that like as one AI agent does something,

17:39 I know that all the files are actually can contained to that AI agent.

17:42 But, there's actually ways that you can actually

17:44 communicate useful information to your other AI agents.

17:48 Very similar to how you operate in a team, right?

17:50 The engineering team needs to communicate to me,

17:52 the marketing team, on how to actually market the product.

17:55 And there's actually ways that we can get AI agents to do this.

17:58 So, that's the next few questions I'm going to be answering.

18:01 Uh, but that's kind of what I wanted to share in this video.

18:03 Narrow agents that run in the cloud, I believe are going to win.

18:07 I think that's what people are going to find the most use from, and that's

18:11 what I'll be talking about over the next few months to run our company.

18:14 I'll see you here for the next video.