Scaling up agentic work with T3 Code

0:00 Hello everybody. Today I want to talk a bit about how you can scale up your agentic work thanks to T3 code. Things move very fast in this AI era. So maybe by the time you watch this, there will be probably other solutions, but hopefully at least this gives you an idea of how to handle multiple AI tasks, working in parallel, letting agents do the work and help you with the context switch that ites without losing track of everything. So first of all, a bit of context. AI is an amazing tool, right?

0:36 So I use it for pretty much everything. a lot of people focuses on the engineering side because that's where a lot of disruption has come but I use it for way beyond that. I spent a lot of my day between engineering hands-on work and management and thanks to the AI I've been able to explode and go back to doing more hands-on work while keeping up with my management duties. That said, I still drive it. I'm still engineering and I'm a little bit still in the loop.

1:09 Sure, there are things that I just let it go like I have a task that has been maybe running for a week but on the dayto-day on the task I want it's a tool that empowers me. Then the issue that happens is when you want to scale the app and that's what I want to talk about but maybe so you can see yourself and know where you are let me talk a bit about the journey right so world started just chatting with chat GPD and being amazed and how smart it was like it talked like a human but also was very stupid right didn't have access to many things and it hallucinated a lot of stuff but things got better we got tools and now it's when people started integrating it into their daily usage with the ID, right? So, if that's where you are, cool. That's not bad. But be careful because the AI solutions that are integrated into the IDs may be a bit dumb. It depends. If the IDE it's using your hardness, for example, if you love cloud code, if your ID has a sidebar that opens cloud code directly in that sidebar, then I think you're good. If you like it, keep doing that. But be careful because not all IDs do that. But just keep that in mind. Even if you're in this situation, keeping with the normal UX on ID and just having a side chat, it doesn't really help. So my recommendation is what a lot of us have gone through. Like the next step was to get out of the ID, consider them a thing of the past and start working with dedicated terminal harnesses. So that's cloud code, that's codeex, that's open code and many others. The point of that is that you are driving now directly from the agent from the AI side. You can still open the code if you want and do whatever you want with it, right? That doesn't change. Like you can even have like side by side your your terminal and your code editor. That's perfectly fine.

3:02 If you are still at that stage, all good. But having it in the terminal unlocks a lot of things like so that's why in the past few months terminals had a huge comeback. Everybody was working on terminal on like improving the multiplexers and all that stuff, right? It allows you to use any harness you want and still getting access to the same code without relying on your ID actually doing anything about it. But even with the power of terminals and I know a lot of people who loves terminals like there is many people that have been working only in the terminal for ages.

3:32 I'm not one of those sorry but I prefer a better UX for a lot of things. Sure, if you are a wizard of the keyboard, you can do magic things with the terminal. But I think at the point that we are, I much rather prefer to have a better UX that help me combat this issue. Scaling beyond a single threat of work is very taxing. It's very taxing for the brain if you are a bit on top of it. Like if you are one of those that is like I just prompt, forget about it and I will see it on the product whenever it's released next week. I'm not there. I still need to know what's happening, right? Even if for some things I don't care about how it's happening, I still want to know what is happening, right? So for that sort of mentality, scaling even just two threads, it doesn't have to be 100, a thousand, I'm like, I'm super productive doing thousand things at the same time.

4:25 No, I'm not talking about that. Just scaling to two threads is already stressful. Like depends on your mind. Brains are very different, right? But the more threats you put, the more stress you put on your brain, it's more taxing. You need to context switch more often. And also, even if you again, if you are not on top of everything, you still lose track of what what things you were doing. Like if you start like one thing, you go to the other, the other finishes or what was which one was that?

4:56 Like it's it's a bit again, it's a bit taxing on the brain. I know that terminals have like multiplexers. I've tried even CMAX which I recommend if you still want to go on the terminal I recommend. Now we have new developments like maybe super logical helps a little bit although that's you know for a different reason but maybe the UX is it's better and it changes things all of it helps but I still need a better UX. The thing started changing when OpenAI introduced Codex which right now it's just a chat GPT app with a different tab that you select. A lot of the things that I will mention may sound dumb if you have not suffered through them, right? But for me, the first step forward in terms of UX was having the sidebar. Like the first time I saw that was in codeex and since then I'm in love with with that app. So the projects made things work in a way better structure. I could see immediately the projects that I that I had open and the threads inside those projects that I was actively working on that right so I think for me that was a huge step but not not only that like the fact that you have best-in-class remote support to be honest they were the first one to do it actually properly you have this third panel where you have extra tools you have source editors you have explorer reviews you can have web browser in there site chats It's basically what they were saying when they tried to make the mega app.

6:27 Codex is this. It's a mega app. It's sort of like the new era of ideas, not focused on the code, but on giving you all the tools necessary to do what you need, right? So, in that case, it's it's very similar, but it's just a different approach. But Codex, it's for OpenAI. Sure, it's like the harness itself is open source, but the app is not. And even though you can hack around the the application, I've done a lot of a lot of custom mods to that run locally that you know hack into the app, but it's a pain because they updated very often. You need to keep keep track of things and the fact that you are patching things on the release, it's not ideal. Also, the everything they do is design for their models. It's designed for how they want or how they see the agentic world working. Sure they are very open about it. So like you can actually put other other models in it and it's it's supposed to work fine but you are going around the corner right also not everybody has access to codeex model especially at work for some reason people and enterprises are preferring to pay anthropic instead of open code. So like this also give you a bit of a struggle if you want to use the app with other things that that are allowed on your on your workspace. enter T3 code.

7:48 So T3 code for me was the salvation. It's the opensource, they call it the control panel for coding agents, right? We're going to go through all of these things and I'm going to tell you how this actually unlocks the potential to go through more. So the fifth thing is that it's open source. It means that the team behind develops it very rapidly. Like you can see some numbers here. It's actually quite impressive how much stuff they are shipping. And these numbers are the are from last week but in the next few days they have even done even more right so it's kind of crazy but the thing that sets it apart is that well you can also contribute to it sure they don't accept everything because they have their vision and like you know they have like hundreds of PRs like nowadays contributing to open source is a bit of a gamble because there is so much PRs and it's so easy in a way to to put an idea and make a PR that you know that friction it's not that anymore, right?

8:44 So like there is a lot of slop and there is a lot of good ideas but how you manage that. So it's very tricky, right? But the team is open to it and and they do a very good job selecting the good parts and keeping a vision intact. But that doesn't mean that you cannot do things for yourself. Like for example, what I have here is actually my own custom version of D3 code. I'm not calling it a fork because I'm not exposing it so everybody uses it, right?

9:09 It's just the same update every day. I keep updating it with nightly nightly changes. It's just that I add a few things on top, right? That's why you see the orange color. That's so I differentiate if I'm running my version or theirs. It's quite useful, but it means that it's very easy for me to adapt it to my own workflow to fill the gaps and the needs that I have. For example, one thing that helped me a lot because I still want to trigger reviews in the way I want them to do, this is all demo data. So, doesn't really matter what you see here, but for example, I have this custom review, this PR button here. When I trigger it, it links to the data of the PR and it triggers my custom skill, right? So, like if I trigger this, easy peasy. I don't have to do this manually. I don't even have to write the name of the skill manually anymore or even write like I have the whole context of the PR here and if I want to review it, I can review it. It was very easy because they they already have this functionality here. So I just added custom functionality to actually put a custom prompt in it. So that's very useful. And like this I have other customizations that are working very well for me. If anybody is interested I may do another video. U if you want to get that just leave a comment down. But let's move to the key aspect that sets T3 code apart. And it's something that a lot of people one doesn't understand and when they understand then they are like so what's the point of T3 code bro this is the point of T3 code for me this is such a huge step versus what all the others are are trying to do T3 code is not a harness it doesn't replace cloud code it doesn't replace codeex it doesn't replace open code by whatever you are using it's an UI it's an application that sits on top of the harnesses it adds a lot of things. It's not just the UI, it's a lot of things, but it doesn't replace the harness. And that's a key aspect because what that means is that you can go and very easily start using it. You don't have to change your workflow. You don't have to do anything.

11:15 You can try right now. Wait for the video to finish, but you can try right now. You can just install D3 code and if you have cloud code, it's going to pick it up automatically. No setup necessary. You just need to go here. You just go to the providers. You have a list. You enable it and it picks it up automatically. Look, I can disable open code and I can enable open code. Takes a second. It says boom, authenticated, done. Because open code is already installed and it just uses that. So that is huge. It doesn't stop there. It lets you have multiple providers, even multiple providers with the same harness. As you can see on my setup, I actually have four providers active, but two of them are cloud code instances.

11:54 It's very easy to set up the same provider multiple times. You can click here and you have all the ones that are available. Some of them are still more more work and some of them are coming soon. If you want one of them really I'm pretty sure there are PRs already. So if you know where to find them, I'm sure they are here and they recently added this nice flow to add them which is actually nice. That that's the other thing. They keep improving in every way. It's crazy how how good it gets so fast and the UI it's actually very sleek. So what does this mean? That you can have TC code with your favorite harness. If you love Antropic, stay with cloud code. If you prefer open source, use open code or pi, whatever you want.

12:37 But it's all on the same application. Furthermore, it means that you can have very easy multiple accounts set up with different harnesses. You can have your personal account, you can have your work account, you can have mix and match, you can do whatever you want because again the code is not a hardness. And for me that's the key aspect of it. It made switching between accounts between work and personal trivial. And that's something that yeah sure you can have an alias on the terminal but then you always you know you have that friction that you always need to restart the process open a new terminal. It's it you know it was not perfect. That's what I'm saying. You can do all of this staying on the terminal. But there is so much beauty and improvements on the UX if you start using T3. So that's the first thing that already you can see how it allows you to do more in the same place.

13:28 But the differentiator was when yeah I think it was Theo when Theo one of the authors of T3 code was so tired about Codex especially in the beginning where Codex was not very performant. It has very rough edges. You can see his videos about it. He was so tired that he said it. I'm going to build my own. And that's how T3 go started, right? So one of the big ideas he had was to use this new iteration on the sidebar. Instead of organizing it with projects and then projects have tasks inside, what he did instead was organize it with a concept of task. You still have the projects and the tasks and the or threads still belong in the project. That's all the same thing. But the interface itself treats them as tasks. So you can see here go I have these conversations but they have extra information right like you you see here that this woke you can you see that I can settle them right so when I'm done with the task I can settle the task and it goes here usually it's hidden you know hide it and it doesn't bother you anymore but it's there if you want to actually continue the work check again it goes it starts working and you see the information of the the of the status working here. There is monitoring if they are observing some process or or things like that. When they finish they they say done. So all of that is integrated on this sidebar which it's key because the sidebar it's what allows you to change context to keep things in mind to to remember where you were what task was doing what. So what is the walk? always is crucial, right? Because if you have some tasks that you are doing that you do on the same thread and you do from time to time, they are recurring, right? But you don't want them to clutter your sidebar. You don't want to clutter your your mental context. So what you do, you can just snooze them just like a task, just like a reminder. Just come back this evening and now it goes here in 8 hours. It's going to come back and you can you can collapse it and it's not there anymore.

15:40 It doesn't bother you, right? But even more the tasks the threads when you work on it and you open a BR for example what I did here I just just do some something and open a BR right so it open a BR the DC code it's smart enough to link this created BR you see to the task so you can see this thing here that it's green with a BR this is a link between the task on D3 code and the BR this means that if I merge this from here from GitHub it doesn't matter if the PR is merged the task it's considered settle and it goes automatically and it disappears from your site so that's an automation that works very well but not only that like if you pay attention to the sidebar you can see there are different icons like it's everything designed to help you differentiate where things are running I know that this is running on my other computer I know this is running on my computer. I know this project is my website. This is T3 code. I even know which agent is doing things because I can start something here and say check last draft and I can say use codeex, use Luna on medium and go ahead and now I see that this is running with codeex.

17:00 It's very easy to again keep the context in your head. This was so good that if you open now codeex chat GPD there is a priority view that it replicates a lot of this functionality. It I think this is still better but that shows you how good of an idea in terms of UX this was that even the open AI team openly got inspired by it. The other small but very important thing is like sometimes like doesn't it happen that you start and you start writing some prompt but you get distracted and it's like oh this one finished let me check what happened and now depends on the app you're using it's very hard to remember that you were doing that and it's very hard to come back to it. Usually what happens is that when you start a new thread the if the app is good enough it remembers it like I think codeex does that like it remembers that you had that written on the same input field and it restores it for you but what this code does is just again treats things as tasks as reminders as notes even what do you have here this is just a draft this indicates that this task hasn't been submitted yet that it's here in a draft mode you know so that's also it's just a small thing but it it helps you remove the friction dayto day. So that's huge. But now even if with this my brain already could scale like nowhere be before like I could do multiple task. I could keep track of them. They come back and I know where I was. That's perfect. But what didn't scale was my machine. I was not able to be on meetings when the agents were doing like five things at the same time. That is no way. So that's why the next key point from D code is crucial.

18:47 It's crucial. So they have like a server client architecture, right? That means that you can run difficult as a server which connects to your agents and it offers of all the functionality. And then the UI, the application that we are seeing here, it's just a front end that connects to that server. This means that you can actually have multiple servers on multiple machines and you can orchestrate them all from the same application on your machine. So as I was showing here, this is running on my other machine while this is running on my machine. You can see Alex M4 and you can see keyless pro.

19:23 You can even have the same project on both machines and you can switch between them. I think there is even an automatic mode where it just sends the work to the the machine that has more free resources at the time which it's I mean I don't need that because I don't have that many machines but if you have a whole fleet of machines this is huge. So not only that but you can open T3 code in just a website because you know as much as it pains me it's just a web app like and it's very good one to be honest. So you can use the desktop app which is an electron app, the web version from any browser, but even more they have mobile apps. So you can connect T3 connect which is a service they provide. It's basically just some layer on top of Cloudflare. So don't freak out. It's nothing creepy. And thanks to this, you can connect to your T3 code servers from anywhere in the world, which means you can take your phone and have the same experience, the same conversations that you can do from your home, from outside.

20:25 If you don't like to use their service, you also can just use Tailscale or if you know how to do it yourself, you can do it yourself. There is nothing. It's all open. There is no no secret sauce here. But to be honest, I'm quite happy with T3 Connect. It works very well and it's very easy to set up. So what this means is that now I can start heavy work on the other machine and I know that while I'm working on my MacBook I can just do my work. It doesn't it won't disrupt me. Like I can be in meetings I can be sharing the screen with the CPU or the GPU at 100 but because I'm sharing the screen and not because the agents are going rogue behind the scenes. So these are the key aspects for me. This has unlocked so much potential.

21:07 The sidebar makes my brain have way less friction. The serpent architecture makes my machine have less burden. And the whole thing about the multiple providers means that I can just use whatever I want and whatever it's best for the task. But it doesn't stop there, right? The a small detail that I think codex gets wrong but difficult got very well is the fact that every project it's sort of its entity, right? And then that project lives in different remotes like as I show you here when I start and I go to T3 code for example I can select which remote this task is running on. On Godex it's sort of like every remote is the top level entity and you need to go it's treated as a separate project which is very confusing. It pollutes a lot the sidebar. So I don't think that's the best experience but the way it's done on T3 code is huge. You can even see how that works because if you go to the project settings, you can see that it detects that this project D3 code is on different machines with different defaults, different configurations. You can even it's brilliant when you change a configuration and it detects that that configuration is not the same in other machines. It actually suggests you just click here and we're going to sync it for you, which is topnotch. Like I love that. But what else? There is the famous side panel like just like in code just like in most of the other UI applications you have this third surface where you like so on the left you have your tasks your threads in the middle is where you have the main thing conversations but then you have this side panel what can this side panel do it can have a browser TC code also has browser capabilities in it so your agent can actually drive the browser and take snapshots and do whatever it needs with a browser to actually iterate and get to the goal that you want. It has a terminal. So this is actually quite cool because with a terminal it means that even if you don't want to do your engineering work with D3 code the fact that you can go into a terminal from here in another computer. If you see here now I'm connected to my other computer. I didn't have to manually SSH.

23:24 I don't have to do crazy setup, which for people that are comfortable with this, it's may not may not a big deal. But for the ones that are not that comfortable, this is another way to control your separate machines from your own computer remotely without having to do any crazy setup. So for me, this select the terminal, it's a huge win for remote setups. But what else? We also have files. So you have a file explorer here. If you open, you have an editor.

23:51 So for people that still wants to manage the code to do to tweak things after the AI has done them or if you just want to do manually, you still have the possibility here. If I go here, I can also show you that it has a div view. So it shows you the div that you have done that it's not committed. The one you did on the last turn. I actually customized it to for my workflow. So I had implemented for myself the fact that I can see unstaged work because that's how I like to work. Like if I'm reviewing something, I start with unstage and I just keep staging it, you know, and it disappears from the list.

24:31 So it's sort of like a checklist that I need to review and then I click reviewing and it keeps going away. So that's also very good. As we saw before, it has pull request viewer. So you can actually go to this button here that shows the pull requests that you are assigned to. So I can go for example on TC code for example and let me change this filter and in a second it will show you the pull request from T3 code and you can just manage it from here in the same way that you do from GitHub. You see all the content of the pull request. You see the timeline of changes that happen. You actually even see the code. You actually can leave a review from here. It's huge.

25:16 You also have those extra things that I was showing you before. It's actually like some days I don't even open GitHub. It's it's that good. And what else? You can see the soup agents which nothing is running here. The side chat, this is also customization that I implemented. I I use that all the time in Cex because I still use Codex for a lot of things because I love Codex to be honest. But this is not at least today it's not on T3 code but there was a a pull request that started the work iterated on that preview a few things and now I have it ready on my fork and now as just so you can see how fast this moves just couple of days ago they added this devices functionality which it it's for iOS simulators and Android emulators and you can actually control the the simulators from here you can even control them from a remote machine. So you can actually have your work being done in a Linux machine and let the agent drive an emulator or simulator in a Mac OS machine for example for iOS. It's not only that. You can actually start the the simulator here and it's going to stream a video to D3 code that you can actually manipulate. You can click you or you can just sit down and look at your agent do it for you, right? But it works. So these are all the extra things that you have on T3 code. That's what makes this also like a mega app, you know, like you can feel that this is like the ID but for the new generation.

26:48 There is a future orchestration V2 that I'm very excited about because one of the key things on CEX is that Codex knows how to manipulate Codex and that's huge. It knows how to start threads, manage threads, delete them, talk between them. It's huge and that's something I'm really missing on D3 code which hopefully it's there soon. This is what works for me. This is what has allowed me to just go from one thing focused at the same time to multiple things and losing focus to now being able to comfortably run multiple things and going back to them and doing more work thanks to the agents and thanks to D3 code. I think this is a huge step forward and I can see where we're going from here.