A clear plan lets parallel agents implement code

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 32:59

   This, which is awesome, it's on social media, on threads, and on X, but can you tell us how you use today Cloud Code in terms of, you know, parallelism, and tips and tricks that you and the team have kind of learned and share across the team? Yeah, I mean, look, there's no one right way to use Cloud Code, so I can share some tips and things, but I think the wrong conclusion to draw would be to just copy these and use it.

2. 33:24

   the way we build cloud code is we build it to be hackable because we know every engineer's workflow is different there's no one way to do things there's no two engineers that have the same workflow it's just every every engineer is the same with workstation setup right like keyboards monitor placement all that ever has it differently yeah it's like we're like crafts people right like you choose you choose your tools like we care deeply about it so there's no one right way to do it so for me the way that i do it generally is i have five terminal tabs Each one of them has a checkout of their repository.

3. 33:55

   So it's five parallel checkouts. And usually I'll kind of round robin and start quad code in each one. Almost every time I start in plan mode. So that's like shift tab twice in the terminal. And I also overflow as I run out of tabs because there's only so many terminal tabs. I used to use web a lot for this, like quad.ai slash code. That's the place that I overflow to. Nowadays, I actually use the desktop app.

4. 34:19

   It's more convenient. it. So quad code, you know, it's been in our desktop app for, you know, for many months. It's just the code tab in the, in the quad app. And I actually really like it because it has built in a work tree support. So that's existed for a while. And that's quite nice for parallelism. So you have multiple, you don't need multiple checkouts. You just have one and then we automatically set up get work trees for you. So you get this kind of environment isolation. The reason I do that is I actually just really hate fiddling with get work trees on the command line, because it's kind of fiddly, like you need to know the cd.

5. 34:50

   And the git work tree, for those who are not as familiar with it, it's when you can check out, instead of having a separate local folder, it's almost like checks out a separate branch, right? And then you can work on it separately, but not have the conflicts only at like merge time. That's right. Imagine that you have a folder, but you have maybe like, git makes five copies of that folder in a way that's very cheap and kind of easy to throw away, so you get this kind of isolation.

6. 35:16

   it can work in parallel and the quads don't interfere. Yeah, so you now have support for this, which I think you recently added like native support, but like for your workflow, you just stuck with the old one of checking out all the separate folders, right? Yeah, exactly. I actually find that over time, I'm using the desktop app more and more for this just because I don't need these separate checkouts and I just have a bunch of quads running in parallel and I don't have to think about it.

7. 35:39

   The other surprise hit is the iOS app for me. Every day I start, like I wake up and I just start a few agents on my phone. Oh, the native one, yeah. The native one, yeah. It's just like, it's the quad app, it's the code tab in the quad app, and it's the same exact quad code. Except it runs in the cloud, right? It runs in the cloud. Yeah, so you have to kind of configure the environment. Luckily, our environment's pretty simple, so, you know, and we just use hooks for it.

8. 36:01

   So you just use the session start hook and configure it. This is kind of one of the benefits of making quad code really hackable, is it's very easy to do this kind of configuration. And this is something, honestly, I would never have predicted, because you know like i i code on a computer if you told me six months ago i'd be writing i don't know a third i haven't told the data maybe like a third half something like this of my code on a phone that's crazy but that's that's what i'm doing today and you're using parallel agents at what point did you start using them and how has it changed your work because one thing that i noticed on myself i don't really use that many parallel agents i may be like two at a time but i'm someone who well i i like to be in charge and especially with claude claude is is a tool that

9. 36:48

   you can follow it along it tells you what it's doing it you can also have for example learn mode which this was shipped a lot earlier where you can actually follow along it gives you tasks i i feel that like Staying in one tab and following along, the model is pretty fast as well. I can kind of keep in touch. I'm assuming at some point you must have done this, but then what happened when you changed to parallel? And do you feel you're losing any control or it doesn't really matter that much?

10. 37:13

    Yeah, I think there's kind of like two modes to think about or kind of like two kind of workflows to think about. So when you're new to a code base, learn mode is awesome. Highly recommend it. For people that are onboarding to the QuadCode team, people that onboard to Anthropic, the thing that we recommend is, so you do, for people that haven't tried it, you do slash config in QuadCode, you pick the output style and you can do learn or explanatory.

11. 37:37

    We usually recommend explanatory because that tends to be better for new code bases that you kind of haven't been in before. For me, once you're familiar with the code base, you just want to be productive, right? Like you just want to ship as much as you can and you want to kind of be effective doing that. So the world really switches. I don't really go deep into tasks anymore. I start a quad in plan mode.

12. 37:58

    I'll have it kick something off. With Opus 4.5, I think it got there. With 4.6, it just really, really does it. Once there is a good plan, it will one-shot the implementation almost every time. So the most important thing is to go back and forth a little bit to get the plan right. So what I do is I start one, I enter plan mode, I give it a prompt. As it's chugging along, I'll go to my second tap and I'll start the second quad, also in plan mode, get it chugging along, then go to the third tap, go to the fourth one.

13. 38:25

    Then maybe I'll go back to the first one when I get notified that it's done. And then I'll kind of... Do you have notifications on or do you turn them off? I actually operate in both modes. Sometimes I do like, you know, focus mode on the Mac. So I just have it off. But also, sometimes I use the system notifications. And you're very, very productive with PRs. I mean, I think it was very visible. Even around the holiday breaks on social media, you actually were responding to, I think someone reported a bug or a feature request.

14. 38:54

    I'm not sure which one it was. And then an hour or two later, it was done because you did it. You've also talked about like number of pull requests you've done on a day, not to like show up, but just as context. What does a pull request typically involve in terms of complexity? Are these like, are some super trivial or some actually like larger pieces of work as well? Yeah, each one varies a lot. Sometimes it's a few lines, sometimes it's a few hundred or a few thousand lines.

15. 39:21

    They're all just very, very different. It's changed so much. Like back when I was at Instagram, I think I was one of the top two, maybe top three most productive engineers at Instagram just by volume of code written. Oh, wow. So I've always, you know, for me, I've always just coded a lot. like this is a coding is like a way that I can express myself. And it's just like it's a way that my brain thinks also. And so now I just get to do it. But I think with quad code, the kind of code that you write, if you are very productive, it tends to be even it's just the number of PR sort of undersells what's happening. Because I think people that used to be very productive in the old days before AI assistants, a lot of the code maybe was like code migrations or something like this so like people that ship you know 20 30 prs every day a lot of it was like pretty you

16. 40:07

    know like a one-liner or kind of migrating a to b or whatever nowadays i ship you know 20 30 prs every day but every pr is just completely different some of them are thousands of lines some of them are hundreds some of them are dozens of them are one-liners it's none of these are kind of code migrations because actually claude just does those and i don't need to be part of that shipping this much code or this much more productive the obvious question that comes up for any i guess software