Claude Code Subagents 2.0 Just Dropped (Forked Subagents Are Insane)

So, look at this guys. This is forked sub agents in action right here. Now, in this video, we're going to take a complete look at forked sub agents and how they're changing the game. GBT 5.5 is about to release as well, but for now, I'm just going to take a look at this from Claude. Let's jump into things. Okay, so at first glance, Fort sub agents probably seem like they're going to be more expensive. The reason being is they actually get more information. So basically the way fought sub agents work is they are fed the entire history of the conversation so far plus a prompt right whereas as of now as of today they are not fed the entire history of the conversation right they're just given an instruction they go and read all the code bit by bit and then they start to generate the code changes right but with four sub agents now I wouldn't have understood this two days ago But because I built Harbor Build recently, I fully understand caching. So the whole point of this is that the read here when it reads the previous or the conversation so far, it is a cached read and not a full read and a full write. This actually makes it easier for Claude code to deal with because it's a cached read prompt which is which is extremely cost-effective and also doesn't use that much power from claude. Now I assume this is something to do with like imprinting or you know the model has a memory of the conversation that has already happened. So therefore it's more lightweight and more convenient and more efficient to read it a second time instead of when it's fresh and new information. Okay. So this is the result. This might not seem like a lot to you guys, but this is very, very close to how I want my thumbnails to be. Whereas recently, I've been struggling to get my thumbnails to where they want to be. And I'm going to show you how that was actually possible using forked sub agents. So, I actually did something that honestly guys, this is one of the best tips I can give anyone and everyone that uses clawed code. I don't know how many people know about this to be honest with you. So, I'm just going to do new tab here. And one problem I have, right, is that I get kind of stuck in conversations with Claude. Like this is my conversation that I'm having with Claude code, right? But what happens if I don't know, I I I go through this entire conversation, I change my skill or whatever and I change the workflow and then I don't actually like what's happened, right? How do I and then let's say I've run a clear. How do I get back to this point? Well, you can actually do something. You can do a claw- r and you can look for conversations in the past that had the workflow or the skill the way that you need it or the way that you feel is better than the current way. Right? Obviously, there's other ways to do this. GitHub, whatever. If you're a smart developer, you probably have backups for all your skills and workflows. If you're like me and you're just a layman using these tools, trying to understand them, trying to have fun, then you can use dash r and you can actually find old conversations that you've had that might fit your workflow and your skill better. I know that that will help someone a lot out there. So, I just wanted to mention that quickly. Now, to activate this guys, literally all I did was I had to do some other things as well to get Claude 4.7 available, right? I've been having issues with Claude 4.7 and homebrew. Basically, I haven't been able to access 4.7 in the terminal for since it was released basically. Um, but yeah, all you do is just I just said, can you set fort agents can they can now be uh enabled on external builds by setting claude code fork sub agent equals 1 and then it just asked me a question. I said persistent and then can you try and activate it? Right? And then it wasn't working in the clawed code desktop app. So what I did was I just hopped on over to the terminal. I finally got access to 4.7 and then I activated them. So, just before we continue with the video, guys, huge shout out from our sponsor, me. This is harbor seo.ai. And I really, really want to highlight something here, guys, because this is me being as transparent as I possibly can with the results and the results speaking for themselves. Now, 224 pages published. There's actually a lot more content than this that has been created by Harbor, but basically, this is the content that is published Now, anyone in product management will tell you that if there's an extra step, an extra two step, an extra three steps to doing something, then a lot of people just won't do it, right? So, what published actually means is they've gone to their library and they've clicked on the article that has been generated and they have linked it to the actual live URL, which allows us to track, right? But even so, even with that in place, right, and it being a little bit too complicated for some people or too much for some people, people just can't be bothered or they just don't do it because they don't want to, whatever. 224 pages published, 70,000 impressions, 720 clicks. This content ranks on Google. I'm super confident in that. We've just updated the writer to make it even better. And you can go get a free trial right now. Harbor SEO is the best AI, GEO, SEO, AEO, whatever tool on the market. And it's also ridiculously cheap, right? I I literally cannot get cheaper than this, guys. So, go and check it out. It's well worth the price. It's an amazing tool. It's got scans, it's got writers, it's got keyword tools, it's got everything. Go and check it out today. There's a link with a trial in the description and in the pin comment. Let's jump back into the video. Okay, so just going back to this presentation here. What happens before is because it was not given the entire history of the conversation. It was not a cached right. It was a full write, which actually ends up being more expensive, right? Which is super interesting to me because this might actually make Harbor Build slightly cheaper for me as well if I can work this out and if it's available in the agent uh ADK as well. So what this also does is it increases the likelihood that the sub agent is going to give you what you want because it has the entire history of everything that you have said to Claude up until this point in the conversation. Right? So obviously this needs to be taken not with a grain of salt, but you need to be careful with how you use this because if you send too much information here, like if you start a forked sub agent 800,000 tokens in, you're going to regret it, right? So what you should do first is you should do a clear, right? So, slashcle and then have a conversation, maybe make a plan, whatever it is. And then you send the forked sub agents to do this. And they receive all of the plan the conversation that you had to get the plan. So, it knows what's important to you, what you're actually trying to achieve, and it ends up being less usage as well. Every time you hit the cache read, it will be less heavy on your usage. Also, just to say guys, this is the GVT 5.5 waiting room. We're all waiting for this, guys. They've built up the hype. I'm super excited to actually see if it's going to live up to the hype. But what I will say is, if you're watching this video today on the 23rd of April, 2026, or if you're watching it tomorrow or in the next couple of days, get the hell off YouTube. Go and get the Codeex app. or if you can't use the Codeex app, go and use the Codeex CLI by downloading it or whatever and go and build what you've always wanted to build. Okay? Nothing anyone can tell you on YouTube is as as remotely useful as just going and building the project while the AI is good. What people are saying is this is their version of Mythos, but they're releasing it because they're so far behind in the AI race that they have to release something that is just probably going to mess everything up. Now whether that's true or whether that's just hype, we're not sure yet. But I'm very, very interested to see. But one thing I have to say is that Anthropic is consistently shipping updates to Claude Code, Claude Desktop, Claude everything, like Claude managed agents. Absolutely insane update. You can literally have the power of Claude code as an agent in the cloud that you can sell to people. That is crazy. So yeah, I have to say guys, this is exactly what I needed from this thumbnail. It's managed to nail it completely and I feel like I have control back over that conversation because of for sub agents and because of the trick that I showed you before with dash r. Okay guys, I'm going to leave the video there because I probably have to go and make a codeex sorry not codeex GT 5.5 video as well. I have to say the updates recently have been crazy. If you've used Opus 4.7 in the last week or so, you will know that although it's not like this mythos level beast, it's a step up from the original Opus 4.6, which when it was released, Opus 4.6 was an absolute machine. So, this is like having that but better. And if you're stuck in thinking, oh, but it's not this insanely good new model, that's not nobody cares about the step between 4.6 and 4.7. 4.6 was already good enough to do anything. episode 4.7 is better than that. What are we even talking about here? Literally a year ago, they couldn't tell you how many hours are in Strawberry. So, just bear that in mind. Stop listening to everyone on the internet so much. Go and build your own projects, guys. Thank you for watching the video. If you are watching all the way to the end of the video, you're an absolute legend. Go and check out harbor seo.ai if you want to support me. If you want to support the channel, there is a free trial. Thanks for watching. Peace out.