He Runs Claude Code and Codex as ONE System (You Can Too)

0:00 Watch this. Every dot is an AI session and they are talking to each other. This is not a demo. Sur my community co-founder is going to tell you everything about it. First, why the

0:10 tools you pick matter way less than you think. Then we will talk about setup and then the short version. What can you immediately implement to get 80% close what this whole thing is doing? You just

0:22 launched it. What we are looking at >> in here? I can see the sessions that I have running and those are mainly cloud code and codec sessions and I can see it when they communicate with one another

0:33 can message common online which sessions are running depending on in which work tree they are in there is a load of engineering happening in the background so maybe let's start with why would

0:44 someone set up their clot code sessions are communicating with codeex sessions and the whole other question is of course the debate between Apple and Android is now debate between entropic

0:58 and open AI. So how do you go about that? Obviously you're using both. >> Yeah, I'm definitely using both. I find that both of them have different strengths. So I may have one daily

1:11 driver. I use clot code the most. I feel the most comfortable with it. I'm most used to how the models work there. But I also see that GPT 5.6 Six soul is a very capable model which has slightly

1:23 different strength than cloth code. Just putting clot code to work, having it do stuff and then having codeex review the work of cloth code is a very effective workflow that catches stuff that cloth

1:34 code on its own wouldn't catch. >> If somebody is choosing between the two, do I read it correctly that your answer would be get both together? It depends how much money you have to throw at the

1:46 AI toys because this is obviously it's not running on 20 bucks subscriptions. >> Yep. No, if you want to get the most out of it, then it will be best to get both because they both build on each other.

1:59 If you're just starting out and tinkering, by all means, just get started with one. I think Codeex is a little bit more generous with usage and it has much more capable cheap models.

2:09 So, if you want to go and use just GBD 5.6 six terra then you're going to get a ton more use out of your subscription. So that's one thing to keep in mind. But I personally like cloud code as a daily

2:20 driver the most. Just opus 5 is like a has its quirks but is a very capable model. >> How do you daily work? Where do you communicate that the models would be

2:30 running the show? >> Right. So this is a view that's cool but I don't really work in this. It just visualizes when the models message each other. So, I built a system, an MCP that

2:40 allows models to discover each other and then send messages back and forth. And this took quite a bit of work to get to work because there's not really an easy process to get this to operate.

2:50 >> I'm curious how it's going to grow because I remember during our calls, you showed me fully working on multiple projects and it's like an army and not even agents. I wouldn't call it agents

3:03 because each of these sessions can have its own skills and agents underneath it and workflows. >> Yeah. >> So this is like harnesses agentic.

3:14 >> I'm not so good at coming up with the terms like wipe coding. I would have never come up with that. >> Like multi- aent orchestration is like the maybe the the easiest is kind of

3:23 what it is. So I really like that my models can communicate with one another. The main reason is sometimes I usually separate them by work tree. So a model gets like a copy of the code base works

3:34 in their own feature and then once it's done merge it back in. This is like the standard way of working that I think every developer who's a little bit deeper into this like knows and

3:43 understands. But sometimes you still want multiple models to work together on the same project. And then sub agents can work but sub agents have their limitations. And one really nice option

3:56 with this MCP service that models can still communicate even though they're working in the same work tree, working on the same project, but still have them understand, hey, there's another one

4:06 working and that one is making commits and I'm also doing work slightly separate, but I'm like working on the UI of the same project and I'm also want to make commits. Let's actually talk to the

4:16 other agent and figure out how we like won't stumble over each other's work. And this is what they can do and allows me to be more hands-off. And one really cool example of that is that you can't

4:26 really do properly with sub agents is opus 5 on max reasoning is a really effective way to do UI work. Like I found that there's like there's no better way. And you can also see this on

4:39 I think the front end arena. You can see that opus 5 on max reasoning outperforms fable 5 at any like level of reasoning and also GBT 5.6 soul. So what I like to do is open up a cloth code session set

4:51 it to opus 5 on max and then open up another cloth code session that set up on fable and the fable model does like the orchestration and I tell it like hey there's a pair available there's a

5:01 different model session available you can discover it with this MCP and that one is responsible for all the UI work and that works really well and then second to that there's also codeex is

5:12 really good at computer use so I tell fable hey you're the orchestrator Opus 5 is here available for UI work and codeex is available whenever you need a model like to do some computer use, adjust

5:25 anything, click on any buttons, that sort of stuff. And that suddenly allows Fable to reach higher levels and be able to delegate work and do more and get stuff to a higher level without me

5:37 having to get in between. And another really basic example, sometimes the context window is filling up, stuff is getting really long, and then you want to hand off stuff from one session to

5:48 another so it has a clean context window and can get going again. And then have the models directly communicating with one another is a much easier way of getting there where like the the second

5:58 model that gets to work that has the clean context window can ask questions to the first one. Hey, where is this? What do I need to do here? And then the first model still has all of the context

6:07 and the second one can then get to work with a clean context window being a little bit smarter because it doesn't have all that baggage from earlier but it can still access the old information

6:17 there as well. >> I love it so much and the best part these orchestrated models can reach out to me and communicate with me and then communicate with my agent. So yeah we

6:28 will talk about the whole involving us as a humans and working all as a team but before that so observability that's the right handle because you can also quickly get in and see what's happening.

6:40 So show me how do you prompt are you still in a terminal or you have interface which you build because I remember that you had also voice interface are you still using that?

6:51 Yeah, a little bit. But codeex being so good with the voice mode overtook that, made it obsolete. No, but I have a single session in here because it's an MCP layer. I've built like a separate

7:03 agent that doesn't need cloud code or doesn't need codec to run that can access and read like what are all the other agents doing and I can just talk to that one. So also from my phone, I

7:13 can reach that one. I can just talk to it like hey, how's this session doing? What's going on with that one? and I can use that remotely from my phone to check in with what what's happening with the

7:22 different sessions and it's a little bit slow but that's perfectly fine because it's very capable. 80% of my communication still goes through the terminal. So we can see here like the

7:32 communication in it in the chat and I'm like sitting here doing this nothing at all and then just wiping each other. >> Don't you orchestrating harness that is orchestrating other harnesses. So that's

7:45 not nothing. That's solution architecting. Huh? Okay. So very practical. If somebody has codeex and clot code or is considering to get both, what is the fastest thing in 20 minutes

7:59 they could put together and do? >> So just having them both and reviewing each other's work is already that gets you to like the 70 80% level of like getting all the benefit out of it which

8:10 I think is very useful to do. have cloud code build and then codeex review is my preferred flow. But I also know that people quite like having codeex work and then cloud code review. I think either

8:21 way works quite well. It's really helpful because it's if you have two different people, they look at stuff from a different angle and the same goes for the models. I also really like every

8:30 now and then just give two different models or harnesses the same task and see how they do, which one you like best because I know that these models have become so capable that it also comes

8:41 down to personal preference. And I know people that really going back and forth with soul much more than going back and forth with Opus, especially with Opus being a little bit quirky with how it

8:53 communicates right now. So I completely understand that. But I think it's important for people to try it out, experience it, and decide on it yourself what you like. But other than that, what

9:02 you can also do is just have a scratch pad that they're both aware of. Putting something in a scratch pad and then using that for them to share information with one another is also very basic, but

9:14 also a quite effective way to have them talk to one another, share with one another without you having to sit there and copy paste stuff from one model to the other. You may have to say

9:24 sometimes, hey, there's something new in the scratch pad. Go check it. But that's much better than having to like copy paste messages in between. >> We're talking cloud codecs. What about

9:34 open source? So I know that in your setup there is open source also running as well. >> I've built my own layer on MRA and I really quite like what they've built,

9:43 but I think it's really important for you to experiment yourself. So if you're all about data privacy and security, then you may want to look more into PI or open code. PI especially if you want

9:54 to tinker and build something completely yourself. But I think it's really important right now to just try and experiment. Decide for yourself what's important. If you for example have some

10:03 sensitive stuff or if you maybe don't want to pay for a $200 subscription over at Open Athropic, then it's really easy to use DeepSeek V4 flash for example inside of Codeex. I

10:16 think it took me about 10 15 minutes to set that up to have it that working. Like I said before, we're getting treated with a lot of different options right now that are really solid. So

10:26 whenever someone said, "Oh, this sucks or this is bad." It may be bad for them >> or the setup >> or the setup might actually suck. They might have a lot of old skills or a lot

10:36 of clotted MD information that's now hurting the performance of the model. But really decide for yourself what you like the most. don't blindly follow like the people saying, "Oh, this is the

10:46 thing like the next Chinese thing you should go for." I think it's much more important to be flexible and to be able to say, "Oh, I'm going to try out a different harness right now. I'm going

10:54 to try a different model right now." >> Of course, this testing cost a lot of money and time. So I also see your your and my job especially with the community on the weekly calls that members expect

11:07 that we are going to toss our money our time to experiment and play that we can direct or give at least informed advice and not just reading articles or PR pieces that what's best or who killed

11:20 what but actually work and put that in the real production and to have a real insight. All right. In our WhatsApp group, it's you, me, and our agents. And basically, anyone can collaborate. Your

11:33 agent messages me. Hey God, I need this. I say my agent to my agent, go figure this out. But then recently, we moved to the bus. In the next video, I want to talk with you about the bus. So very

11:46 curious about the mistakes, observations, and is it the new Slack for all the agents and humans to work