OpenClaw Sub-Agents EXPLAINED (Stop Getting Slop From Your AI)

All right, folks. This a this video is all about sub agents for your open claw and it's probably one of the most vital skills. >> Actually, it is. >> Actually, it is. >> There's a massive improvement in the >> the quality of the work. >> Exactly. And um I show you this in example. Um and you know, knowing me, >> I'm going to build a presentation about sub agents on a video about sub agents. So if you just look at my prompt, so you don't really have to set anything up for this actually. >> Yeah. >> Yeah. So it's like you just tell it to, >> right? >> Well, maybe a bit different for Mini Max cuz uh it kind of forgets the next day to do the same thing for uh using parallel. >> Okay. So we'll we'll talk about that very clearly in a moment. But for for this video, I'm going to keep it simple. We're using the Opus model, so the smarter model. So he knows a little bit better how to deploy sub agents. But let's just take a look at my prompt here. I said, can you make a presentation on sub agents for openclaw? Go send sub agents to research why it's important then come back to me. Use sub agent to make the presentation and uh make it good with other sub aents uh and make it good. Okay, and other sub agents to do the SVG graphics. So this is like a very very very loaded prompt. Okay, it's very loaded. It's like sub agents everywhere. Okay, and this is mostly to test it out and show you guys why it's kind of good. >> All right, so yet again, so Stark here, he says he spawned two research agents. So one he's looking at the web right oh one he's reading the open claw documents and the other he's reading the web >> right and then he came back with a huge load of research so this is one of the reasons why maybe you guys are getting like slop yeah work right if you're not deploying sub agents then you're having the one guy kind of do this stuff and he wants to rush and give you the results back but these agents are now working in parallel there's two agents they're working on results and they might even come back with different answers That's even that's even better because the AI can help you decide and even if it hallucinates yeah if one agent hallucinates the other one can kind of correct it its course right this is why sub agents will drastically reduce how much kind of slop you're getting in your AI work right >> and like this is insanely long this is just from one prompt and you didn't have any prior MD files you tell it to read right >> no no no but that being said because this agent knows me he knows how to search on Twitter he knows how to search. So, he actually gets a lot more information. You can just tell like you can just compare this with like maybe a very simple search. This is just not the same. You're getting much better results here. >> Um the other thing uh that I want to say just fundamentally and I want to make this say this very early on in the video is that um sub agents why they're good is because why Ron? because of the breaking up >> context. >> Context context window optimization. Yes, you drilled this into my head. >> Yes, I was drilling this into you yesterday. I hope you guys know. But uh because the sub agents, they don't have to know everything about you. They don't they don't know they don't need to know, you know, they don't need to know what to do in the morning. They don't need to know how you brush your teeth. They they're they're dumb agents, right? They they um because my orchestrator knows everything about me. I only need him to know everything about me. all these sub agents because they're dumber, they have a much smaller context window and then they're going to perform the function a lot better. And this is very very applicable to someone that's using something like maybe mini claw or miniax um claw or kimmy or the Chinese models. The reason why is that the Chinese models perform a lot better with smaller context, right? So if you want to use a cheaper model, >> please and you can also configure your sub agents to use different models, but I'm not going to talk about that here. I'm just going to use it, you know, on a on a whole. But you can see that because they have a smaller context window, they can come up back and dump the information to my main guy and then the main guy can really make the presentation for me. So that's that's kind of how um the flow works right now. >> Mhm. >> I guess the other thing to point out is that sub agents actually also improved in the last version of openclaw. Yeah. So if there's any failures in the sub agent, >> then it'll tell you now. So before I think sub agents were very wonky. I think like it was like uh >> you know because you you were frustrated that your sub agents failed and just didn't come back with anything, right? >> Yeah. Not even a notification that it failed, >> right? But with the latest version of open claw, the sub agents um actually give you a update. Yeah. >> Um and if if they fail, they'll still, you know, they'll update you. I fail, then you can just run it again, right? So you can see that it made the presentation for me, you know, the the kind of parallelism that we just talked about. M. >> So technically it's just the same thing that we're repeating here. But the parallelism is also very good in sense that >> because it's doing everything in parallel, >> it saves you time. >> Y >> but at the same time because I feel like at least for my agent, he's really pressured to give me results fast. >> So being able to deliver results fast means that um he'll be sloppy if he's not running things in parallel. >> Yeah, >> I guess that that kind of makes sense, right? And I think by default if you get [clears throat] set up with your open claw agent I think by default they don't run sub agents or parallel uh agent workflows unless you specify. >> Yes. Yes. So just specify. So you can test this out by just saying hey look use sub agents use a bunch of sub aents and you'll see the resulting difference. So hopefully if you can run some a command like this see if they give you the same quality of work that these are giving. >> Yeah cuz that's the end game. >> That's the end game. That's the end game. You want a full detailed research, but obviously this is going to cost you more because every agent is going to run in parallel. So, you're doing more work, so it's going to cost you more tokens. >> But that being said, um if you want to run this like more cheaply, uh the Mini Max plan, they give you a lot of um uh it's a lot of um prompts are included in the package, so you don't have to worry too much. If you're an Opus, >> yeah, maybe you should be a little bit worried. M but that being said so I guess the TLDDR is that you're doing more work >> but the work each amount of work that you're doing costs less because of the context window because the sub agents have less context so they just do the work and come back right so that's kind of how you know uh it's a balance and from from our preliminary testing the results are much much better all right so um there are also like um differences like >> uh between single agent multi- aent I guess you have to you have to think like an orchestrator like what kind of tasks get broken down. >> Yeah. >> Um the AI mostly most of the time is smart enough to break it into research agents or I I found that yesterday during my workflow that when building a presentation for me if I just asked my main uh agent to build it he was really bad at it. >> But if I said hey break this task into sub agents. So, one agent do the graphics, one agent do the uh presentation. It actually broke down the task to kind of chunk sized pieces which they managed to do better and perform a lot better on. >> Okay, I'll keep trying that with Mini Max. >> Yeah. Yeah. And then I kind of guess we um went went too hard with this. But um as per as on the rules for this channel. So >> um as part of this channel, we want to deliver you real insights and real work cuz we're using this in our daily work and seeing if you know our life is improving and you know obviously it is. >> But that being said, uh what we're doing is Ron is using Miniax more. I'm using Opus more. And we're going to do that comparison with you because like I feel that's very insightful for you guys um to watch because at the end of the day, you know, what model do you choose? Do you really spend do you really fork that $30 per day to use Opus? >> So far, we're kind of leaning towards my side, huh? >> Yeah. So far, in terms of the quality, of course, your side. >> Yeah, the quality is just a little bit better. So, guys, try out sub agents. I'm not going to make this video too long. I think that's it long enough already. Just try it. It's already built into your open claw and the new version is a lot better. So, if you have problems with it before, you probably won't have it problems with it now, >> right? So, anyways, with that guys, uh, subscribe to this channel. We are growing so quickly. like we are growing at a massive immensive rate. So, thank you guys so much for your support. We also look for every single comment on our videos and we try to update you guys as much as possible. So, make sure you're subscribed and then ask any questions because we make videos for you guys. All right? If not, then I'm just going to take a break, chill out. [laughter] But yeah, with that guys, um yeah, thank you guys so much for this. And then if you have any particular issues and I also want to thank our community as well, No has been particularly good. He's been providing insights that we haven't even noticed before. So, Uh the whole idea of this channel is to just grow with you guys, learn this AI stuff, and hopefully have real applications for you guys to use. So with that guys, thank you guys so much for watching this video. See you guys in the next one. >> Shing out.