T3 Code Orchestrator V2: Multi-Agent Workflows

0:00 I'm in Codex, but the work it's actually going to be done by Open Code, all thanks to T3 Code. Here we have it, sent by another agent. And here we have Open Code delegating to Codex and Cloud Code to use an adversarial review of my latest blog post. Hello. If you're new here, T3 Code allows you to manage multiple agentic harnesses into the same UI application. In the last post, I went into many details about that, but today I want to show you the new capabilities that are available in T3 Code thanks to the newly merged orchestrator V2.

0:42 And remember, a model generates responses, but a harness is what provides the tools. And T3 Code works with different harnesses. It's not a harness itself. I keep repeating that because it's so crucial and people get it so wrong. For example, here we are using Open Code, and we have many different models because Open Code is a harness that harnesses all those models together, right? So, we can use it we can ask it to analyze the latest post using a sub agent with another model. It's still on the same harness. So, what Open Code is going to do is going to execute its own sub agent capabilities to do the task that we have given.

1:21 And here we can see Open Code triggering its own sub agent capabilities to analyze and do the work that I've asked for. But now, what if you could do that across harnesses? That's what T3 Code now allows you to do, and it's what I've been waiting for such a long time. Why? Because you can do things like this. We can use T3 delegation to ask Cloud Code Opus and Codex Saul to critique my latest post. So, what this is going to do is it's not going to use the harness itself.

1:55 It's not going to use in this case Open Code, right? What it's going to do is it's going to use the tools injected by T3 code itself. And these tools, especially the delegate tool, allow any harness that you're running through T3 code to spawn sub agents in different harnesses, completely different harnesses, not in the running one. So here you can see how we have a sub thread owned by Codex. So this is now Codex. You see GPT sole ask answering for the question that open code asked for. If we go up, which you can navigate through these right panel, we also have the we have cloud code doing the same with Opus. Thanks to this, we can have one agent that it's our main driver, like in this case I was using open code, but then use the benefits of all the other agents and all the other harnesses, in this case Codex and cloud code, to use their models and their full capabilities. And all of this is interconnected thanks to T3 code, and you can bring everything together into the same interface, and just delegate naturally. Orchestrating across harnesses is so so much simpler thanks to this. Opus was actually faster, and it finished sooner. And now it's giving back automatically. We didn't have to copy paste, we don't have to actually delegate manually anything. Once the sub agent finished, even if it was running in a totally completely different harness, it brought back all the results to the main agent.

3:27 And now we can see how Codex finished with sole, and again, it brought back all his feedback. It seems like I'll have a lot of work to do, right? But the capabilities of T3 code as an orchestrator don't stop there, because they took a page from Codex, a page that it's like one of the best features that Codex has. Now we also have them available in T3 code. So let me use my Cloud Code harness here, and I'm going to say start three different D3 code threads.

4:02 Each one to brainstorm a different idea for a blog post. Use Open Code bunny for one, Cloud Code octopus for the other, and Codex so for the third. So, with this, we are now not just using sub agents, right? So, like as we saw here, this is a sub agent that runs inside the same thread, okay? This is like the typical sub agents that you have in all the harnesses, to be honest, like Cloud Code, Codex, but the advantage here is that it crosses through the harnesses, and all the communication is handled by Cloud Code. But here, what you just saw was a thread that I asked Cloud Code, in this case, to create D3 code native threads that spawned and very nicely animated here for each of the things I asked. So, each of these threads, you can see at the bottom here that it has a different harnesses running the brainstorming session that I wanted. So, this unlocks a lot of potential because you can not only delegate the work to a sub agent, but you can actually create threads and do the work yourself.

5:31 I'm still hearing gates. So, now I have three different things that I can engage here and start answering and having different conversations that are all persistent. They're all native D3 code threads for tasks. I can settle them, I I reschedule them, all from a prompt in one thread. Not only that, but it can also manage the thread itself, not only create them. >> >> I'm done with the Codex one. Settle it.

6:01 Reschedule the other two for tomorrow. So, everything that you can do through the interface yourself as a human, T3 code now can do with your tasks itself. So, it's giving these superpowers. You can see them animating. They have now scheduled or been settled, and I didn't have to do anything. This is a tool that unlocks a lot of potential, and now you can build here just your workflows, use the skills on top of these to actually make it trivial to manage and orchestrate not only multiple tasks, but also multiple tasks across different harnesses using the best thing of them.

6:41 And finally, I just want to show you how these orchestrator B2 capabilities and features are not only available from T3 code itself, but T3 code exposes an MCP server that you can add to your favorite harness. So, for example, in my case, I use a lot of T3 code as my main driver, but for certain things I still go back to Codex. And now it's trivial to jump back to T3 code because as I showed you in the beginning, you can have with this MCP, you can ask Codex to spawn a thread in T3 code, and it just did. You see? This was a spawn from another agent. It means that this comes from this Codex functionality. So, you can navigate back and forth between the harnesses, and everything is managed by T3 code. This is Honestly, it's another unlock on the potential that you have with this tool.

7:39 So, this is what now T3 code allows you to do. You should try it. Right now, as of recording of this video, these capabilities are still in nightly because they still need to polish them. But honestly, I recommend you to change your update channel to nightly because it just works. I never had any problem with using nightly. And once you have access to it, I recommend you to start simple, right? In the same way that now you're asking your harness to delegate some work, you can actually, instead of using the harness sub agents, if you have multiple subscriptions or you're using different harnesses, you can actually benefit from that. You can actually ask the delegation to happen across these harnesses. Investigations done by Opus, but implemented by Soul, or you can use some open model in open code, but then for the heavy lifting, you can just delegate to Astra for something. You can go back and forth, even, you know, use a new Haiku or Luna, if you prefer, for the small things. You can configure that in your agents.md.

8:38 You can configure that as a skills. And now, it's going to be like your agents understand which one is better for its task, and it doesn't matter who's providing them. It works across all the harnesses. And that's all for today. I hope you learned something and you enjoy using T3code to actually do your work because it's such a beautiful tool. That's all for me. Thank you for watching and see you next time.