Orchestra — coordinating 10+ coding agents without becoming the bottleneck

Speaker 0: Formation layers for managing multiple coding agents, which, is really topical right now. We're all thinking about how to manage multiple agents. This should be an interesting 1. Now we're dropping on the Zoom. I promise next time, I'm gonna work we're gonna work out Zoom. Speaker 0: It's gonna be seamless transitions all the way through. Alright. Yeah. You got that mic's a little thing there. Testing. Speaker 0: You Speaker 1: guys hear Speaker 0: me? Yep. Sweet. Speaker 1: Alright. Thank you so much for taking the time, everyone. My name is Paulo. I'm cofounder and CTO of a company called The Hog, backed by Gary Tan and Y Combinator. And, I've been using coding agents for the past few years, very aggressively. Speaker 1: And 1 thing, especially once CheeseBack came out a few weeks ago, or a month ago, I'd say now is that Paulo of the planning and, the execution and all of those skills are really useful. My number 1 concern, I'd say, still when I'm working with coding agents, especially in parallel work streams, is keeping track of everything, the context, the decisions across all the layers. And that part is really challenging. I've tried a lot of different solutions before. I tried using, you know, Git as always, and I I still use that Hubly. Speaker 1: And I would always just basically research, implement, plan, or plan and then implement and write all the progress down in my, markdown files locally. But the 1 thing that became apparent, especially once I started using Gstack to implement, like, very large end to end features, is that oftentimes while I'm working on a, on a specific feature and then I go out and work on another parallel feature, at during the implementation or the planning phase, I might make an architectural design decision or engineering decision, and then that doesn't translate to the other more trees that I'm working on. And so basically sharing all this context across everything is really frustrating, and so that's why I built out Orchestra. And it's a very simple, like, I tried not to over engineer at all. It's all locally, or it runs on your computer completely locally. Speaker 1: And it's basically just a file system structure. And the way it works is that, every single time you, so here, let me I actually prepared a small demo for you guys. So first I I cloned the cal.com repo, so it's the open source project for the, cal.com. And I actually found here a few issues on their GitHub. So I found here a bug on, their Google Calendar integration, another 1 with Google Calendar events. Speaker 1: And the main thing that I wanted to point out here is that all all of these issues here are booking and scheduling, related bugs. And so the first thing that Speaker 0: I did is that I already Speaker 1: did a prompt here. And is that I'll go over to my coding agent. So I'll go over to Claude, and I'll just copy paste this prompt. And the idea is that this prompt will just, go out, you know, try to understand all the issues that, I I want to research. And, you know, I might do this for a feature, but in this case, it's just fixing several bugs that are touching, similar parts of the system. Speaker 1: And the first thing that it's going to do is break it out into, a first thing Speaker 0: to send a problem and Speaker 1: then break it out into several different plans. And then once it does that, you can see that it automatically picked up it automatically loaded the orchestration skill in internally. I'm gonna cancel that just because it's gonna take, like, 7 minutes, so we don't have that time. But, basically, it'll go out, document all of its research, and document it out into a thread here, in Orchestra. So this is just saved at the root directory of my laptop. Speaker 1: And so you see here it says, thread 13 booking reliability. And it basically, first has, like, a conversation MD file that documents what the user, aka me, is talking to the agent about. And then it'll also, keep track of a a spec. And so this keeps track of all of the related parts of the system that it might need to touch. And then, inside of this, inside of this session, I basically, started creating a plan to actually fix these things. Speaker 1: And so I created several different, PRs here, like, a plan for those PRs. And then I actually have a progress file. And, basically, when I have multiple coding agents, working on this, it'll basically go ahead and update itself. So, let me try something here Speaker 0: for the live demo. Speaker 1: Let me blow the mic down once again. So the first thing I'm going to do here is that this is actually my Jira workflow, so I'm just gonna go in here and say, hey. Give me a prompt for each of these PRs so I can implement them parallelly. And, also another thing to mention is that for Orchestra, I created this decisions directory. And basically every single time you're working on a thread, it'll actually have the context of all of the engineering and architectural decisions that you made. Speaker 1: So, every single time the, a new session spins up, there's something called the session hook. And so, it's I just used a bunch of pod code, sessions for, like, start session, stop session, compacting sessions. And, basically, these hooks allow you to to in, to inject the Orpstra context back into the, comics window. And so I have all of the cumulative decisions in here. And so once I have all of these, these prompts here, for example, so, like, fix this, this issue right here, for example. Speaker 1: I can just go copy paste this, and then I'll open up, I'm just gonna use superset here just to make it easier. And I'll just copy paste this and then copy paste the second 1 as well. So these are prompts that we'll just, go ahead and work on the plan that we, created with Work Truck earlier. Let me create 1 more. This last 1 here. Speaker 1: So this is a web look, issue on the GitHub issues. There we go. Yes. I trust this folder. There we go. Speaker 1: So now all of my coding agents are working in parallel, and they're all first checking into Orchestration. And they're looking at all of the, all all of the threads in here. So they're first looking into this high level thread, which is all the com related, conversations that I had. So first, it's trying to understand, okay, how did this user, first kick off the conversation, then it's looking at the plan and the current progress. And then as it continues to work, they'll actually use sub agents to update the progress of the plan. Speaker 1: So we should see in a bit here that the progress will change from to do to, done or being implemented. But we don't have to wait for that. I just want to point that out. So, yeah, that's pretty much it. And, yeah, that's basically how I've been managing to to work on, like, 5 different PRs at any given time, throughout the day. Speaker 1: So Yeah. Speaker 0: That's really great. It's great to see lots of people working on this prop right now, and this is a really interesting approach. So, have we got any questions for the audience? Yep. Is there a public git repo? Speaker 0: Because there's a lot of information. I know why does that Just to repeat the question, asking if there's a public git Speaker 1: repo. Yes. A 100%.