Multiplayer Coding Agents in the Cloud - Charlie Holtz, Conductor

0:00 Building anything great takes a bunch of people and I so I think in the future it's going to take a bunch of humans and a bunch of agents to build great things. >> Let's go ahead and bring on our first guest, our only guest today. No stranger to the show, has been on before and I'm sure a lot of you watching have probably used Conductor. So why don't we bring on Charlie from Conductor. Charlie, how's it going?

0:21 >> Good, good. Thanks for having me again. It's good to see you guys. >> Good to see you, too. >> Honestly, I don't remember how long it's been. It would have been at least three or four months, I think, since you've been on. And so I know you've been cooking a lot of things up over there. >> Definitely been a lifetime in in AI time. It's funny, I was thinking about last time we talked was also definitely pre-Fable. And one of the main things that has changed for us is that we don't use our local computers anymore. We don't use local hosts, we don't use work trees. Everything just happens on a remote computer that sits in the cloud somewhere. And it's kind of crazy to think that I used to live in a world where like work trees mattered so much.

1:01 And so yeah, it's partly because the models have got It's actually almost entirely just because the models have gotten so much better that we've had to adjust to this new cloud world. So yeah, excited to talk about how we thought about cloud and how we got there and but also just yeah, good to see you guys as always. >> Yeah, I mean I think one of the things that you were definitely one of the first proponents of was don't look at the code. You know, you were you were definitely early to that and it was you know, people were wondering, "Well, where's the IDE? Where's the editor?"

1:31 And I don't know if you ever even added that. And you I know you could like view the code in there, but I there's probably a lot of pressure early to add that. But you I know you were very resistant to it because you kind of saw in the future a little bit and you're like, "There will be a time when looking at the code will become less relevant." Not saying, you know, some folks listening, I still occasionally look at code, but it is much rarer than it used to be.

1:53 >> Yeah. It's funny, we anytime we're thinking about adding a feature that either internally we want or user wants, we try and think of like is it going to be ahead of ahead of the park or like where the park is or a little behind it. And one thing that we really debated was adding like the terminal in the bottom right. When we first released the app, which is now a year ago, I really didn't want to do a terminal because I just felt like that wasn't the future. But we still have the terminal and so that's one thing that I actually think has like remained useful for a while. In terms of like viewing the code and editing the code, we did finally acquiesce to user demand and we added a thing that we call caveman mode where you can edit the code. We're using a Pierre's diffs library for that, which has been amazing.

2:39 Shout out to the Pierre team. They're like they're awesome to work with. But yeah, we also don't look at the code that much anymore and it's like steadily trended downwards as the as the models have gotten better. I think like there are occasions where like the looking at the diffs like the diff viewer is important just to get like a sanity check. Honestly, actually the main reason I use the diff viewer is to like see which lines that AIs have commented on. But yeah, that even even doing that is trending downwards for us.

3:06 >> I mean the growth of how the product has changed so quickly is has been so awesome to see. Also, congrats on your series A by the way. >> Thank you. Thank you. Thank you. >> So when did this shift for you? It's almost like conductor became cloud agent native overnight. Like it just came out of nowhere. But it obviously was seeded for a while. So like at what point did you know where you needed to shift and like how did you both retain your current user base on the way there?

3:32 >> So we've been talking about this since December. That's That's I remember when we had our first conversations about cloud becoming the next important thing. I think we the reason we like hesitated and went back and forth for it so long was because we just weren't sure how long the models were going to run for. Like the thing that I just like kind of like made it clear that we needed to do this was when Fable came out, it was like oh crap, like these are going to be these are quickly we're going to quickly going to be in a world where these can run for like not just like minutes, but hours, days, like even weeks. And obviously like if something is running for weeks, you can't just let it run on your you know, my I have a MacBook Air like sitting right here. I don't want to just like leave it propped open for a while. But the other thing that's really important is like when we launched Conductor a year ago, it was crazy at the time to imagine running like even three to five cloud codes at the same time.

4:26 But now that the models are better and you can let them run for longer, it means that you can also have like more of them running at the same time. Because you don't need to like constantly be context switching, like oh what is what feature is this one working on? Does it need me to prod it? Oh, is this one finished plan mode? Do I need to like approve the plan? Like you can trust the agents to do a lot more work, especially with like Fable class of models. And so because you can run so many more now, it's even crazier to like imagine them constrained to the computer. And then we started experimenting with multiplayer ideas.

4:57 The gist of that is like I want to see what everyone on the team is working on at the same time. And so once we started like seeing the value there, which was now like probably two months ago, it was like clear that cloud we had to go all in on cloud. And then in terms of like bringing our users along for the ride, that's a that's an open question. We're still like figuring out how to like there was a couple of issues, like you you're basically getting a brand new computer. And so setting up a new computer always sucks. No one likes doing it. And so like part of our job now is to figure out how to make setting up a new computer fun and easy. Actually maybe first first step is easy, then we can get to fun. But it's hard.

5:34 >> You're exactly right. I think we're going to look back at the time of last year when people had their laptops cracked open, right? All the time like like walking around like this. And then then there were apps, a lot of people had like apps that would just keep it running so you could close your screen and and that was a big thing for a while. And we're going to look back and kind of laugh at that because you're exactly right where you know, whether I was running cloud code, master code, conductor, if you have too many sessions opened, it's going to bring your computer to a halt, right? And that's I think everyone's probably experienced it at some time where they just are trying to run too much and their computer slows down. Why should you be constrained to just your local machine, especially when you pretty much want to kick off this task, let it run for a while, and then you'll check in on it later.

6:16 >> Yeah. And and maybe I actually I could share my screen and show you. >> Yeah, we we love seeing demos. So, if you can show a demo >> I wanted to give a quick demo of something that I think is everyone is overlooking in cloud right now, and that is multiplayer. The first like instant thing that you get from moving to a cloud world is that like this is a this is a cloud workspace in conductor.

6:34 as far as the user is concerned, it looks and feels just like a a local get work tree, but you can see this little subtle green cloud icon in the top left and click it and see that that actually is running on the cloud and you can see the little information about the computer you're you're using. but everything on the right also just feels exactly like normal conductor. Like I can click a file and view the file changes or even like make edits, and it just streams that information from the cloud workspace to my local conductor app.

7:04 So, like that's like the stuff that everyone like already knows is cool. Like I can quit the app and the agents are going to keep running. I can have a bunch of them all running. It's not using up a bunch a ton of RAM. But the thing that people aren't talking enough about is when you are cloud first, you can see what everyone is working on in real time, and then you can actually like prompt together, which is pretty cool. So, if I go to this homepage, I can see these are all the workspaces like across all my projects and and then I can see these little avatars, and those are people on the team. So, I can click on this workspace, and it's going to it takes a second to load, and I can see that here on the top right that like I'm online, Cayden's online. I can see his prompts. and if I scroll to the bottom here, I can actually see in real time that like his agent is working.

7:55 and I can see what he prompted, and if I want, I could be like, "Hi Cayden." I don't actually think I should mess up his workspace cuz he's probably It looks like he's doing real work. There's a a small like a few kind of cool UX things that we can do here. So, like we can tag people, and I can be like, " Jackson, like what do you think of this?" I can click up here and follow the workspace. So, like even though like I I didn't create it, I can like have it in this little following section in the the top left. But also, we've given all our agents access to a conductor API. So, you can see here that the agent actually decided to create a workspace itself.

8:33 And so, I can click that, and it will take me to a workspace that an agent decided to create. and here at the bottom, I can see the workspace it was sent from. And so, like all of this this stuff theoretically could work in a local world if you like always had your computer open and like caffeinated it or something. and then like we you uploaded your teammates' transcripts to the cloud somehow. But anyway, it would be it would be really confusing. And so, by just going cloud first, it makes like multiplayer super seamless. It makes the ability for the agents to create workspaces super seamless. And then obviously, we now have an iPhone app, which is going to be too hard for you to see, but I'll go to this workspace and I'll just say like, "Hi from iPhone."

9:24 and send it. and then you can see the message just popped up. >> Sick. >> Yeah, I mean, I think the the whole concept of multiplayer wouldn't work without cloud, at least not intuitively. You You want to be able to stream the results from, you know, so you can I can see the same stream you're seeing. We can steer the agent together. >> Exactly, yeah. >> The one thing that this does, which is an unlock that has always been something difficult to do, is like hand off work to someone else.

9:50 >> Yeah. >> I mean, you can know, of course, you just push up a PR. It's not that challenging, but it's even better if you can actually see everything that went into creating that work. What was the person thinking? How were they prompting it? You know, what were they explaining? Maybe you can look back at their at what they were saying and find flaws in their logic instead of just flaws in the code. Like actually flaws that I wouldn't have done it this way. I think it's going to end up in the wrong spot because we didn't provide the right details on whatever the feature is, right?

10:18 >> Yeah, and the session context, right? The agent has the context from that session, even if I'm jumping in and taking it over, as opposed to what you have to deep research the PR when you take it over, and then start from scratch. >> And and you're still guessing what the original context was, right? It's like it's lossy. >> The way I think about it is like code is almost like the sawdust of the conversation. It's sort of like the byproduct that you could throw away in 2 months when a new model comes out, and then you can just like replay the conversation again and get new code that's better, magically, because the models are better. but the other analogy we often talk about is like Figma made it so that like in one click, you can see in real time what your teammates are working on and like see an artifact of their work. and so our goal is to make it so instead of creating a PR and then like sending that, and then going to Slack and copying and pasting it, and then opening that in your cloud code, like opening the PR or like copying and pasting parts of the messages, says again, one click, I can press command shift C. I don't even have to click, and I have a link to my workspace that anyone on the team can access. And that has felt really magical.

11:23 >> So, another good like implication of this is like I mean, there's a loneliness problem in SF, but there's also a loneliness problem in engineering teams because no one pair programs anymore, right? There's no sense of camaraderie when you're doing your craft. How have you seen that in your own team? Are you pair planning? Are you pair Are people collaborating again? >> People are collaborating again. I think our the most common use case has still been review. That's mostly when we've been using it. It's just like I have a workspace, I have all the changes, and I just want like a quick sanity check on something. But it's a good point. I think I think even we even though like we're we're at the frontier or pretty close to it, we're still kind of figuring out the new like cultural norms. Like do we at each other in the workspace and then like collaborate like that or do we like prepare a bunch of plans and then share that?

12:14 >> These new norms are being invented. We haven't quite figured it out yet. >> I mean a weird byproduct, which I'm not predicting, but if I was bold enough, I might predict is, you know, conductor for your team basically becomes like a a second slack. I could squint and see that that being a possibility where, you know, here's the feature we're working on. Let's just collaborate here with the agent watching and interacting alongside of us. I don't know if all your chats are going to go into conductor, but I could certainly see a subset of communication ending up there for sure.

12:44 >> Yeah, especially engineering discussions. And I think if if anyone is watching this works at one of the labs or is working on models, interesting problem we run into is that all of the language models are trained for these like one-on-one turn conversations where like I say something, then Opus says something, then I say something, and Opus says something. And like Opus might say a lot of stuff. Like it might have multiple turns within its response. But it's not trained for a world where, you know, I say something, and then Abby says something, and Shane says something, and then the model responds. I think this is partly like a harness problem, but also partly just like the way the models are trained. And I think that building anything great takes a bunch of people.

13:27 And I so I think in the future is going to take a bunch of humans and a bunch of agents to build great things and so I I think there's opportunity to train the models to understand what a multi-person turn multi-turn conversation looks like. >> We struggle with this too with our agents. Like we built some like we built some primitives in Mastra so you can have these multiplayer conversations, but that doesn't mean the model is good at responding to them.

13:52 Like we can queue messages, we can insert context, but it still like has to acknowledge that there's other questions that are like not addressed, but it's still like so singularly focused on the thing it was working on. As humans we can do multitasking, multi-context in a single conversation. I could respond to both of you at the same time with threads or whatever, you know, and but it's agents aren't that way. >> And very importantly, you know when to not say anything.

14:19 >> Exactly. >> I'll say like at Abby just FYI like you should be aware of this and you'll know that you don't need to respond. Whereas the models love to they love tokens. They love to >> Thank you for that, Charlie. Thank you so much. >> Like thank you, yeah, exactly. It's like thumbs up, like looks great. And then like 500 words on like its thoughts, yeah. >> I'm curious Charlie, why you did why did you decide to do the iOS app versus like a common service people have is Slack, right? Would you have been able to achieve the same thing just using the Slack app or did you want a rich mobile experience?

14:52 >> Great question. So we do have a Slack internal app that will release as like a first-party thing. But something I didn't mention that I also think is super cool is that we have a API now. So anyone can build anything they want on top of Conductor Workspaces. And so we've seen people build their own Slack bots and their own like mobile apps and even one person made like this like chicken farm game where like the chickens are Workspaces and they're like eating when they're working and then they like like look up when they're done. But, long story short, Slack is too noisy. I think there's like a lot of exciting interface ideas that we have, and so we just wanted to like control the interface for mobile. Like live activities while the workspace is working, or push notifications, or like seeing the avatars when someone's in the same workspace, and like things like that.

15:41 >> Plus, you get to control the design experience, as well. >> One of the things we've said on this show, if you if you were having watching for a while, for those of you tuning in, Avi and I always talk about, you know, either the bar test or, you know, sometimes we call it the beach test, right? Is what can you, you know, kick off and then go and finish the work from the bar or from the beach, right? And and I think we're getting closer to that being possible, right? With, you know, with the mobile app, with, you know, there's there are tools you can use within Slack or whatever, but we can basically take it with you and actually be very productive just from a mobile device.

16:14 And you don't have to bring your computer everywhere. >> Yeah. >> I think there's a world where the mobile Conductor mobile app is the primary way people are using Conductor. And then I think there's like a world ahead of that where it's agents who are the primary users of Conductor. They're the ones spinning up workspaces, deciding when something is finished, deciding when to merge, and stuff like that. >> All right. Well, anything else you wanted to show off today, Charlie?

16:37 Otherwise, we appreciate you coming on the show. I think a lot of people probably know what Conductor is, but if someone is living under a rock and they don't know what Conductor, they haven't heard of it, where do they go to get started? >> they should go to conductor.build. Thanks for having me, as always, and it's it's crazy how much has changed in like 3-4 months. Next time I'm on, I'm curious what kind of world we'll be living in. but yeah, thanks as always.

17:01 >> WeMo cars, and, you know, just coding from our phones. Charlie, are you guys hiring? >> Yes, we are hiring. if you live in SF or want to move to SF and are excited about building the future of orchestrating agents, reach out to me at charlie@conductor.build. >> Every week in AI, something insane happens. >> And there's so much drama. Every Monday, we break it down live. >> We do the news. We bring on guests building in the space.

17:27 >> And we go deep into the stuff that actually matters. >> Agent Saur, every Monday, noon Pacific. >> Follow. Don't miss it. >> Peace. >> Peace.