Rails Business
Brendan Buckingham and Ryan Frisch talk about developing with Ruby on Rails and how to leverage it to build a business.
Rails Business
Scaling AI-Assisted Development
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Today we discuss lessons from our previous episode with John Nunemaker, focusing on the Conductor tool and how multi-workspace, multi-agent workflows can speed up coding and bug fixing. Brendan shares his experimenting with parallelizing six Rollbar error fixes at once by generating prompts and running them in separate Conductor workspaces. They compare this with a one-agent-at-a-time workflow and debate whether closed-loop setups (ports, databases, dependencies) could make parallel work trustworthy and reduce context-switching costs. They also cover using full Honeybadger/XML reports for faster debugging, using Claude Code from a phone to create PRs, challenges syncing session history, and broader AI product strategy like APIs/MCP, and RAG.
00:00 Catch Up And Recap
00:43 Conductor Workflow Overview
02:40 Parallel Rollbar Fixes
04:35 Manual Testing And Setup Hurdles
06:33 Debugging With Full Reports
08:46 Ports Docker And Dependencies
11:50 Parallelism Versus Focus
14:57 Closed Loop Trust And Context
18:24 Merge Conflicts And Acceptance Gaps
22:16 Review Bottleneck And Output Surge
25:26 Fixing Gallery Uploads
26:39 Multi Select Challenges
28:27 Branching Without Conductor
31:28 Claude Mobile Workflow
32:34 Session Sync Friction
34:44 AI Brain For SaaS
38:01 APIs And MCP Table Stakes
41:49 Internal AI Assistants
43:46 Access And Safety Concerns
46:31 Second Brain Revival
49:03 RAG Tooling Experiments
50:04 Wrap Up And Listener Feedback
LINKS
Questions or comments, email us at railsbusinesspod@gmail.com
all right. So I know, uh, you know, we talked to John Nunemaker last week about some, about a lot of stuff, but we focused on some AI things that he was working on. Uh, and I think both of us sort of left that call, you know, perhaps in a state of like, okay, there's a lot more that we can be. Uh, doing, and I know we both sort of jumped in, but you had some stuff you wanted to share. So the floor is yours, sir.
Brendan:Yeah. Yeah, it was an interesting conversation. That one's actually live now, so, um, people should be listening to it right now. I'd be interested if we get any feedback. But yeah, I mean, the biggest takeaway for me was the conductor tool that he introduced us to, and, you know, I, I probably had heard him talk about it, but I hadn't paid any attention to what it actually was. But it's, it's pretty cool. I don't want to go completely into what it is because John. Pretty much described it in the last episode, but to summarize it here again, in case you didn't listen to that episode, you know, it basically just takes. I would say 99% of the workflow for at least the development side of, you know, opening a branch, testing it locally, you know, all that stuff, Claude Code, pull requests, pretty much everything, in my opinion. It's like everything, but the ticket itself is integrated into the system and for me, I've been for several months now trying to figure out like ways to. Do code reviews. I don't write a lot of code, but honestly, I've been trying now to write more code, or at least have Claude write more code in smaller batches because of something like conductor. Where the biggest unlock for me is like the workspaces idea, and I mentioned in that episode as well that I've been using Claude desktop for a little while. Because that also integrates with Claude Code and it, it alone has the work tree idea and GI, where you can essentially spin up a new repo every time and then you don't have to worry about conflicts. And I mean you do with merge conflicts, but it's not on your local, it's, you know, you're not gonna get a branch conflict 'cause you got too many branches or whatever. Um,
Ryan:It's
Brendan:it's been really.
Ryan:of the server and the database, not the, not the GI repository,
Brendan:Right. Exactly. Yeah. Because otherwise you can only have one branch on your repo at a time and Yeah, exactly.
Ryan:Well,
Brendan:then there's a bunch of others.
Ryan:but it gets confused,
Brendan:Yeah, right, right.
Ryan:it to do whatever you want, but if it starts competing with itself, all bets
Brendan:Yeah, exactly right. And I don't want Claude having to figure that stuff out. Um, but yeah, so that, you know, for me, where I took it was I, I wanted to push it. I don't wanna push it to the limits where like it's gonna break or whatever, but I just wanna push it to what can it do if I ask it to do like six? We use Rollbar, you know, as our, our error monitoring tool. And most of the time we're not catching up and we're not keeping up. Even with some of that stuff, we, we monitor it. Obviously if there's like a, a major red flag, we're going to get to it right away. But there's like Sidekiq worker errors that happen a lot that don't impact customers. They're not really impacting anything, but they're there. We should probably at least put a guard clause in, right? It's like, it seems simple, but I decided to go, like the whole, uh, process, I guess workflow of I was already having Claude. Cowork, I guess it is. So a different piece of claw or a different clawed version. Going out to Rollbar, finding like the top 10, not, not just top 10. I guess it was like top 10 worst offenders in Rollbar and I was having a report back, but I hadn't been acting on it. So now that I was aware of this conductor thing and obviously already using Claude Code. I decided to say, give me Claude Cowork. Give me a prompt for each one of these that I can hand off to Claude Code. And then I took each one of those, and this, this was manual right Now. I'm sure there's a way to automate it, but at the moment it was, take those, I think I did six to start six prompts, put 'em into six different conductor workspaces and let it go. So I essentially had six branches being worked on at once. And I will just make it clear that I didn't even get to the point where like John was talking about getting the actual workspaces up and running where you got different ports and all that stuff. I didn't even do that because these were such small things. I just pushed 'em up to our staging environment to test them out.
Ryan:Right.
Brendan:But
Ryan:You didn't care that it couldn't spin up a server to actually test its work. You just took the fix that it proposed and then tested it yourself. Off
Brendan:I was doing what I could do to get it going as fast as possible, but before I can roll it out to the rest of my team, yeah, we gotta circle back and get that other stuff working. Um, so I I, I'm gonna stop there for a minute. That's kind of the workflow that I've done. It's been, it interesting because it's been good in a lot of ways, but it's also. You know, it still has some limitations. So any other thoughts or have you been messing around with that at all? I guess before we go any further?
Ryan:Yeah, so I installed a conductor. I used it to fix a bug.
Brendan:Mm-hmm.
Ryan:then I haven't had a chance to get back to it.'cause I need to set it, I I need to do the setup up scripts. Right.
Brendan:Yeah, right.
Ryan:think you brought up an interesting point, which is, you know, my workflow to date has not, so like I, I'm using um, you know, Claude Code vs. Code extension,
Brendan:Yep.
Ryan:right. So it's got access to it does a pretty good job of managing branches. And occasionally it runs the test suite like
Brendan:Yeah.
Ryan:It's like, oh, I'm gonna see if the test pass. So obviously it can boot the server, right?
Brendan:Right.
Ryan:But the majority of the work to date never actually runs anything. Right? It proposes a solution. It creates the pr. I look at the pr, I manually test the PR if I need to,
Brendan:Yeah.
Ryan:and, and that alone for me has been, I mean, I'm. Ha. Hard to say exactly how much more productive, but definitely more productive. In fact, my partner, uh, g chatted me today and is like, I'm excited to see how much more you're pushing.
Brendan:Hmm.
Ryan:but it is really there. There's one of the, well, two other things I wanna mention. One is this, this notion of using it to debug. So John had mentioned he just grabs the XML export from Honeybadger and puts it into quad code, and I was like. I never thought to do that. Like usually I, you know, copy what I think the relevant parts are and um, you know, give it a prompt. But I just started paced with no other prompt. Like, I
Brendan:Yeah.
Ryan:help me fix this bug. I just put the bug report in, and so far it's a hundred percent now. So, you know, I, I haven't gotten any like really nasty bugs yet to, to try on this one. But even the ones that I spent. I'll say hours trying to figure out solved in five minutes.
Brendan:Nice.
Ryan:Right.
Brendan:That's efficiency. I.
Ryan:yeah. Well, it, it really is. And it's like, oh, I didn't think to look at that. Right. And, but once I saw it, I was like, yeah, that's it. Right. So like there's still this notion of, you know, quality control, right? Like.
Brendan:Totally.
Ryan:This is why I don't think, you know, software engineers are going anywhere, right? There's still
Brendan:Mm.
Ryan:field that you can look at something and be like, yeah, that's directionally right or directionally wrong. Um, and, you know, maybe AI will get it to a point where, you know, we get comfortable with, uh, where they're at. I mean, I, I think reality is AI is probably just fine and it doesn't, at least me, it hasn't gone astray in a while, which I, I find fascinating.
Brendan:Yep.
Ryan:So I started doing that. I realized how efficient that was, and then I immediately, you know, this week I hit the bottleneck of I am sitting here waiting for Claude Code to finish so I can do the next thing,
Brendan:Yeah.
Ryan:right? And that is the conductor thing. So tomorrow I'm hoping to sit down and get, you know, the setup script working, or at least try to figure it out. But. Yeah, that, that's sort of the next unlock for me. So I'm curious, like you said you hadn't gotten into any of the setup stuff yet, or have you looked at it and just didn't have time to finish it or where, where are you at in that journey?
Brendan:Yeah, I, so I, I did start it. One of the first prompts I actually did with in Conductor was, you know. Try and do the setup script, but at at the same time, while I was doing that, that's when I kicked off those six Rollbar jobs. So I'm at a point where like my local environment isn't usually set up or when it does, it quickly goes, gets broken. And I don't gen generally maintain it because I'm not writing enough code or I hadn't been writing enough code. But I do have it to a point where it's like. It's pretty close. I've actually got the port thing working. So what John described by having, you know, I forget the ranges, but, but basically each workspace gets its own port and then its own database. I had the port working, but my databases are all kind of mucked up. That's a whole thing. I'm not even gonna get into why that's the case. It's a thing with my. Uh, set up and nothing with Conductor. So I think that the concept that John laid out there would work, but my biggest question at the moment is like, we recently have Dockerized, our app, and I know your app is Dockerized. So, and I know John was talking about using Foreman, so I don't think he was using Docker, but I guess we never asked him specifically. And I'm sure you can make it all work with Docker, but I, I'm still kind of curious. What that would change and how that would work. Um, I'm sure you'll probably, well, actually I don't know. You, you don't use Docker for your app, right? You just have it for like when I use it and stuff?
Ryan:Yeah, so I, we dockerized it so that developed, you know, contractors that we work with can spin it up quickly and not have to worry about all the dependencies, but I, you know, it's the only app I work on.
Brendan:Yep.
Ryan:So having it run locally just removes the overhead of to download and spin up images.
Brendan:Right,
Ryan:Um, but yeah, so I mean, I get a little bit of both. I don't go into a doctor very often unless I'm changing docker config, and then I need to test it to make sure I didn't break anything for everyone else's using it. But,
Brendan:right.
Ryan:um, yeah, I don't, I don't know. I'm, I am concerned about the, the, the port stuff. You know, our, our app defines a port.
Brendan:Right.
Ryan:I don't know why we do that locally. Honestly, it was probably just left over from, from something that doesn't matter anymore,
Brendan:Yeah.
Ryan:that feels like it's gonna be painful. I also don't know how to handle yet. I mean, and I haven't looked into it, so it's not like I, I'm sure we, it's a solvable problem, right? But it's not just, we, we have Redis, we have memc, we have all these other external dependencies.
Brendan:Yep.
Ryan:Um. That just need to be figured out. And I don't know how to, I'm hoping Conductor just has a simple solution to sort of name space, those two. Uh, but if, like, for example, we can't name space Elasticsearch, I mean it's, it's sort of a moot point'cause the test won't pass without it.
Brendan:Yeah, right. Well, so one thought that I've got, I've actually been listening to Justin S's podcast a bunch. Um, I guess I was catching up on some older episodes that I'd missed, and he doesn't use Conductor, but my experience so far with Conductor was cool to be able to run six rollbar errors at a time, but I'm, I'm still not sold. John also was like pushing really hard on if, if he could do a hundred things at once, he'd be doing a hundred things at once and so that's great. I'm glad that he's there. I'm not quite there yet though, like, lemme just take a step back, I guess and I'll get back to what, why I brought up Justin, but where I'm at is it was cool to run him. It was cool to have 'em working. I didn't really need the port stuff and. I, I am, I'm not even done with all those tickets yet. Partially 'cause of other priorities, but it's, it's a lot to have to keep in mind what is going on in six different rollbar errors, let alone whatever else you're working on. And,
Ryan:the whole point of conductor though, at least. I mean, at least from a, having not actually used it yet,
Brendan:yeah, no, I, I think you're right. Like. That's how John's pushing it and that's kind of what they're going for is like the ability to have tons of stuff being worked on at once. And for some people that works. I don't know. I guess where I'm going with it is for you, for instance, with some of those issues you were worried about, I don't necessarily think you have to, to, depending on how you wanna go about it, depending on how efficient it is, you might not have to do multiple. At once, the the, but there would still be a benefit. So the benefit in my mind would be, even if you can't figure out all that port stuff, or it's such a pain, you don't wanna do it, there's still a benefit to being able to have the multiple different workspaces open. So if you have to go back and forth between things, you don't have to worry about the, at least like the work tree stuff and some of those conflicts there. Um, and the reason I brought up Justin s stuff too is because. He was asked, like at the end of his podcast recently, he does a ton of this AI stuff. He's been doing it, I think, longer than both of us. And someone basically said like, what is your workflow? They were just curious, like, seems like he's getting a lot done. What is he doing? And he's actually against having multiple things running at the same time. And I kind of felt. Like, as he described why, which was a, a version of, it's the context switching thing I kind of talked about where he feels that it's better to just kind of get that one thing done. Now what he does is he has one, one thing coding at a time. One agent, I guess, coding at a time. And again, not using Conductor, but he's got one Claude Code coating at a time. Excuse me. But then he's planning the other things with Claude Code so that once that one's done, he can, you know, essentially get another one going. How quickly that process works and everything. He didn't really describe, but uh.
Ryan:Well, I have a few thoughts on that. One is I, I think the, in my mind. It is this notion of, you know, whether it's conductor or another tool, I don't really care, right? But the idea of the ports and the name spaces, and actually giving AI the ability to have a fully closed loop without your,
Brendan:Yeah.
Ryan:um, you know, integrate, you know, without you being part of the process is the thing that. Will allow you to context switch, right? Because you would plan it. You know, you become a prompt engineer, right? You have to plan it. Well, if you don't plan it well, it's gonna produce crap, right?
Brendan:exactly.
Ryan:plan that you agree on, just letting it run and run until it passes the tests or whatever your criteria are completion, like I. I don't believe that. Context switching. I think that's a artifact from how we've always coded that we are uncomfortable letting go of, but ultimately when you have the system set up at to a point where you have a fully closed loop and you trust it.
Brendan:Yep.
Ryan:there is no reason not to have multiple going now. I think the problem is we don't trust it. One, and B, we haven't actually experienced the closed loop system because you said, it doesn't actually boot up the server half the time, right? It just looks at the code makes a change, but it doesn't actually verify the change. So like you have to then come in and verify it, which means context switching is actually problematic. right. So I think that is the hope is that. These closed loop systems that have full access to the actual app running on a port and the databases and everything else that support it be even more powerful to the point where you don't have to worry about contact switching. Um, that, that's one thing that came to mind. The other is, I haven't found like. There are certain tasks that I have to work on that I would only work on one at a time, and that's when I have to load it up into my head. It is complicated. I don't know what it's gonna look like yet. I'm sort of in, you know, experiment mode, right? Like Spike, whatever. Context switching is really expensive there. But for these bugs that are coming from Honeybadger. Or even very small changes, context isn't that large, right? It's like, all right, it's like, uh, I have a, you know, a n error bug, right? It's like, all right, like I see where it is. There's not a lot of context. I need to solve that bug, right? It's a guard clause. It's missing data. It's, you know, a, a bad if branch. There's only a handful of things it can be, and most of the time the fixes. straightforward. I just don't wanna spend the time to figure it out. And turns out AI is really good at doing that for me. So like, you know, I think it is also a matter of big the task is and how much you need to transfer your knowledge. Um, you know, so this isn't like, Hey, I'm always gonna work in parallel, or I'm never gonna work in parallel. I think it really depends on the type of. for that day, it's warranted, sometimes it's not. I mean, gonna kick off a project next week where it afford me the ability to do multiple things at the same time.
Brendan:Yep. Yeah, I mean, that, that makes sense. And I, I definitely agree with a lot of what you said there. I, I think a couple observations I had when doing the multiple things at once that. I guess I'll, I'll, I'll say that I can definitely see getting there, but there I have a little, a couple concerns that I've ran into that I'm sure are just part of the learning curve. Um, but for instance, one of'em was, I gave it six rollbar errors, like I said, and. It ended up two of the rollbar errors were in, it's on, they're on the same page, but they were, they were legitimately different errors that manifested because of different, uh, like one of 'em was a bot basically hitting us, looking for a robots txt file. And that manifested one type of error. And then another one was, uh, something a attribute miss missing. So it was like one of those nil errors you were talking about.
Ryan:Thank you. That was the word I was looking for before.
Brendan:Yeah. The interesting thing was that because I threw it off or, or kicked them off in six different workspaces, those workspaces weren't collaborating and I didn't, because I was just testing it. I didn't take the time to look and say, oh, these are both on the same page. And then it wasn't until I, it opened the pull requests that I realized it was actually changing the same code too. There was like a slight variation in one of the pull requests over the other one, but. And it didn't really break things, but it, and again, also I, I've kind of come to realize after the fact that the way I approached it at that point was I told. I forget which one I, but one of the Claude Threads, I basically said, Hey, this other poll request that you opened, or the, another agent opened whatever, uh, is basically doing the same thing. Can you take a look at that and just like combine them? And it did. But another approach that I probably could have or maybe even should have with the whole closed loop philosophy would be if I had just trusted that that. Would've been caught by like a merge or like, you know, an update of the branch by Claude. Once that first piece was merged, it wouldn't have really mattered anyways. But my old habits were like, oh, we're, we're hitting the same thing, so I don't want to have two pull requests that are doing the same thing. Um,
Ryan:question is like, why, why do you care? Like it's no different than having to merge now. It's just, I think you, I, you identified that AI would be better at working on'em at the same time than you would be.
Brendan:right. Yep.
Ryan:I'm not gonna comment on whether that's a good thing or a bad thing. Like, it may still be practice to look at them in isolation and then let it in a third task, handle the merge.
Brendan:Yep.
Ryan:you know, maybe, maybe that's a bad way of thinking, right? Or, you know, we need to be like, all right, AI is inherently capable of maintaining multiple ideas at the same time, unlike us.
Brendan:Mm-hmm.
Ryan:So why waste the resource? I, you know, I, I don't know.
Brendan:Yeah. Well, the other thing, I'll just add that it was kinda the other part that. I'm still trying to work out, was neither one of those pull requests actually, were looking at like the robots txt file. I mean, obviously it, it did know in the case of the nil error, it, it knew what was going on there, but it didn't really know, or, or I guess when it was testing or like the test it wrote. It, it, in fact, I'm trying to remember if it actually broke when I tested it on staging manually. Like there, it wasn't actually quite done. I think that was the other part of it. And so it felt like there was some sort of a gap where the acceptance criteria that it comes up with are basically run the tests and if they work, we're good. And I know that there's gotta be, I shouldn't say, gotta be, I know there are ways. To make it better, have smoke tests, whatever, you know, system tests that actually test that sort of behavior better. And I didn't have those running, so that could be part of the problem. Um, but yeah, so those were just like observations. They weren't deal breakers. Were now, I'm never gonna do it again. Ultimately what I want to do though, is I wanna understand like what are the, like what are the limitations? What are the, the benefits of this? And if we can do 90% of our rollbar tickets this way now and I can have it set up in a way that I can at least hand it off to one of my developers on my team and not have them have to completely create a new process, that's a huge win right there.'cause then we're keeping up with stuff.
Ryan:I mean, I, I quickly found out this week that, you know, the bottleneck is no longer code, and
Brendan:Yeah.
Ryan:this for a while. The bottleneck is in review. So the question is how do we, how do we gain more trust in the code that AI is producing? Or provide, you know, some sort of guardrails where we trust it enough. Right. It,
Brendan:Yep.
Ryan:coming back to this word trust'cause it feels like those things are converging a, a little bit for me. I mean, maybe it's because I, I'm doing more of it and I'm seeing, you know, consistently better results. Um, you know, I think I pushed. I was looking at this yesterday. I pushed 23 branches this week, so 23 prs.
Brendan:It's a lot.
Ryan:have another 15 open from
Brendan:Wow.
Ryan:Honestly, the majority of that code was written between the hours of seven 30 and nine. Why I am putting my kid to bed,
Brendan:Yeah.
Ryan:my phone, right. Prompting, iterating. Right. So that's great. That's a efficiency hack, I guess. I mean, otherwise I was just sitting in there twiddling my thumbs. But um.
Brendan:Well, let's put a pin on that for a second. Like, keep going, but I wanna put a pin in the whole, from your phone part, but keep going with what you're saying.
Ryan:let's, I, I, I'm just wrapping it up. I think the, I forget what I was saying before now, but the, you know, I, I pushed a lot of code this week and it is now, you know, I, I have more branches open than I have time to go through
Brendan:Right.
Ryan:close, so it's like. Which is great. I've solved a ton of bugs that I had not been able to figure out before,
Brendan:Yep.
Ryan:Customers are happy, my team's happy, right? Like things are moving. Um, but now I'm like, all right, how could I get better results? I, I'm trying to remember what I was gonna say. I was gonna say like, out of all those branches, let's call it 35 or 40 branches that I opened. 85% of 'em were like, good to go on the first cut. Right? And the others are like, I think they're good, but I need to figure out a way to verify. and then there's a, I think there's like one branch that I looked at and I was like, no, this is just, and I, I've tried it four different times now, actually. There's one branch. I'm like, no, this is just not right. Like I, I don't know if I'm prompting it wrong, there's a misunder. I haven't figured out what the problem is, but like I've tried it four different ways I've gotten four different results, none of which were what I was looking for. So, you
Brendan:Outta curiosity on that one. Is it, is it another Honeybadger ticket or is it something else?
Ryan:no, this was definitely, uh, more complex,
Brendan:Okay.
Ryan:situation. The, the, the quick version is. We have a media library that's largely written in jQuery that can be plopped in anywhere where you want to add an image and it pops open a modal and allows you to upload with drag and drop bulk, upload, you know, full editing capabilities, captions, al Tech, all that stuff. And then it drops it into, you know, whatever the placement is on the thing, on the page, you know, whether it's a blog image or share image or you know, website. or whatever. Um, we also have the notion of galleries, right? And because of the limitations of how the media library existed, when we built the gallery feature, like you have to add one image at a time, So a gallery of more than say a, you know, a handful of images is like painfully tedious to. Um, to create. And on top of that, you can't reorder anything, so get it right the first time or forever hold your piece, right? It's not a great UI for the user. So I'm trying to solve both of those problems, which is right. How do I allow, allow multi-select in the media library so that you can sort of. You know, bulk upload, pick all the images, hit enter one time and your gallery's done. And then on top of that, add the reordering component,
Brendan:Yep.
Ryan:which are, are, you know, those are two separate things. The reordering was actually the E, you know that that actually worked. The first cut. Figuring out how to get the media library in a backward compatible way so it doesn't break everywhere in the app that is used to then do this multi-select thing has not, um, I haven't quite gotten there. Like there, I think there's pieces in each of the four branches that are like, okay, that that's how I, I think I would do it. Now let's, I gotta just go back and be like, all right. at these four things. This is what I like about each combine, the good things, and I think we're there, but that's not something I can do from my phone though there there's too much context to copy, paste, communicate. Not to do that from desktop, but. That was the, that was the only one out of the 40 ish branches that, like, I, I feel like I hit a bit of a, I won't even call it a dead end, but like work, like actual work, um, to get it to where I needed it. Everything else was, you know, varying levels of complexity. It's getting better at JavaScript, believe it or not. I, I don't know. Like I've had a real hard time with that. And I would say, like, thing on my wishlist is if it was better at Flex, flex and grid, flex
Brendan:Oh yeah.
Ryan:I not get that. It never gives me what I want. And you know, maybe that's my own limitation on understanding the differences in the language. But, um, yeah. All right. So you wanna talk about. Mobile. What was
Brendan:we even do that, I'm curious, since you haven't been using Conductor a lot or much, it sounds like, how did you get so many pull requests open? Like what? What is your current workflow without conductor?
Ryan:Um. I mean, it, it's, it's one at a time, right? But the challenge is like, I spent more time, I felt like this week just waiting for Claude to finish its task
Brendan:Yep.
Ryan:and immediately being like, I wanna switch branches and work on this other thing, why I'm waiting. But I can't because if I switch the branch, you're gonna lose the context
Brendan:Right.
Ryan:now. When I work on my phone, like when I start a lot of these, like I'll, I'll probably create two or three branches a night, like after I unquote sign off work and I'm with my family, uh, hopefully they don't listen to this. Um, you know, late night sort of coding from my phone, like averaging two to four a night and it's usually like, oh, I saw this honey badger come in. Let's see if I can just. Pop that offer. We have a, a customer request. That doesn't seem too complicated. Let me see if I can get it to pop something out for that
Brendan:Yep.
Ryan:I'll come back. Then, you know, I'll do the work on my phone to a point where I think it is, yeah, this, this looks right. And then in the morning when I sit down on my computer, I pull open the pr, I pull it locally. I. You know, tee it up in the, in the app and I qa it manually like I've always done. And then I'm like, okay, yeah, good burn, push. Um, you know, I run the test suite in between, but I mean, that's my workflow. But I'm, I. I am feeling the need to parallelize the work, right?'cause I'm spending more time waiting, um, than I am coding now, which, you know, is a good problem to have, but it's a very different problem than I've ever had because it usually I was the, the bottleneck with creating the code and now not, not the case.
Brendan:Yeah, it definitely sounds like you would benefit more from conductor than I would even so far. Um, so yeah, I, I think.
Ryan:I don't, I dunno, like I, I haven't had the time to sit down and get it working right or right now, like the only thing I could do with Conductor is use it the same way I use Claude,
Brendan:Yeah.
Ryan:fine, but misses the value completely. I.
Brendan:Yep. So why don't we plan on checking in, see how that's going, you know, maybe next time we record the two of us. But back to the, the remote stuff, I'm, I am curious like, what are you doing there?'cause I know when we recorded with John. We brought that up and then I think Claude was just releasing their remote version of Claude Code or whatever it was, I guess their Claude app. Is that what you're using? Are you using some,
Ryan:That was just really, I missed that.
Brendan:well, I remember when we were recording with John last week, this came up and I think you had said you, you wished you were able to do that. It sounds like you've been doing it a lot or you already were doing that. Maybe I misunderstood.
Ryan:No, it, the mobile app has been around for a while, so I've been doing it for quite some time. The problem that I have is, like, the stuff that I do on the mobile app isn't available in VS code, so the, the conversation history,
Brendan:Oh, so you're just doing like conversations with it on the mobile app?
Ryan:Yeah, I, it lit. I mean, it works exactly the same way as the desktop app,
Brendan:Yep.
Ryan:You go into Claude Code, you say create new session. It spins up, you know, it pulls the repo. It has its own environment. It does all, its
Brendan:Oh, okay. Got it. That's the part I was missing. I hadn't done that. I, I knew the mobile app was there and I'm looking at it right now, so, yeah. I see code. So you're saying you are doing branches on the mobile app, but it's like pulling it down, you know, in the, the cloud, I guess.
Ryan:right?
Brendan:Yeah, yeah. Yeah.
Ryan:it's not actually running on my phone. So I, I, you know, I'm, it's a chat interface and it's doing
Brendan:Yep.
Ryan:and when I get to a point where I think it's done, I say, okay, you know, push to GitHub,
Brendan:Right. Got it. So this is all
Ryan:when I pick it up, I pull from GitHub, I run it locally on my computer and I continue. But the problem is, if it's not done. I can't from my laptop or from VS. Code, see the, the history of that conversation on my laptop, right? So if I want, if I need something to be like, oh, I thought we already addressed this. I have to open my phone, scroll through and find it. Right? So those environments don't share, they don't sync, right?
Brendan:Yeah. And that must be what they just released. I don't think it's live for everybody, but I had not used code, Claude Code on my phone. I knew that the apps, like I knew if I was doing chats, you know, just normal chats, they would sync. Um.
Ryan:Yeah.
Brendan:I hadn't messed around with the Claude Code stuff at all. I thought I misunderstood last week when we were talking that the whole remote thing that they just released in beta or whatever it's in, was going to allow what you were already doing. So that's interesting. That's, that's cool. It does sound like it's, yeah, it's definitely got some kinks to work out, and maybe that's what that remote thing is all about, but,
Ryan:I mean, that's my only complaint is that if, if the sessions would sink so that I could reference them,
Brendan:mm-hmm.
Ryan:be fantastic. Right? Like, I don't have to do it very often, but sometimes I, you know. I'm done for the night and it's like, all right, I'm halfway through and oh, it is what it is. Like, you know, it's not necessarily the worst thing to have to start over the next day in terms of like picking up where you left off, but there's more I could do there to make sure I don't lose the context, but
Brendan:Yeah.
Ryan:hasn't been a big enough problem. It's more of an annoyance at this point.
Brendan:Yeah, that makes sense. Um. Well, cool. I mean, I think that that covers most of what I've done with Conductor and there's a lot of ancillary things and just even like non-coding related AI things that I don't feel like I wanna dive into right now. I, I kind of wanna poke on some more of that stuff and see what that looks like. And maybe we, that's another episode or something, or at least part of another episode, but. So maybe this is a good place to even just stop and I'm gonna continue, continue using Conductor. I know you gotta get it set up, but I would imagine probably by the next time we're recording you'll hopefully have that set up and we probably have more thoughts on it as well. So anything else you wanted to cover?
Ryan:I had one other thing that came out of the conversation with John that I thought would be interesting to talk about,
Brendan:Shoot.
Ryan:you know, he, he had talked about his AI brain for his company.
Brendan:Oh yeah.
Ryan:Um, and that got me thinking about, you know, how do I position LOC bowl to, you know, survive the ai, call it adaptation because I, I don't, I'm not in the, this is gonna kill all SaaS products boat.
Brendan:Yep.
Ryan:But I, um, do think if you refuse to adapt that it will kill you. Right. And I spent a lot of time thinking about that, and I think the challenge at the moment is, you know, envisioning what that future looks like.
Brendan:Yep.
Ryan:It's, it's an unknown. And I, and I'm. I was taking the concept of his AI brain, like how do I feed in all the things that we know? Right? And I think that's a good place to maybe start for internal use of ai, right? But I think there's a concept there that's the same for, you know, our users use of ai. It's like, all right, how do we encapsulate our way of marketing, the local marketing, you know, marketing 3, 4, 5, like all the context and experience we've built up. Over, you know, 15 years and how do we make that information available to make our customers better at what they are trying to accomplish?
Brendan:Right.
Ryan:and I, I think there's something there, right? And I'm starting to get to a point now that I'm feeling this, know, John pointed it out. It's like there was a big jump, you know, call it a 10 x jump. I don't know what it actually is, but a 10 x jump in December when Opus came out. And you know, he's proposing there's another 10 x jump from using something like conductor and multi-agent sort of workflows.
Brendan:Yep.
Ryan:I'm definitely seeing that and it's added sort of fuel to my fire of like, all right, is actually moving faster than I thought. Right. And, and let's just, you know, compare it to the last big jump, which was like, you know, mobile phones and that happened over. You know, a five to six year period perhaps like from, you know, early adopters to laggards.
Brendan:Yep.
Ryan:This is gonna be, you know, if I'm being like worst case, like half that, probably a quarter of that in this. So like you're talking less than two years to massive change in full scale adoption of these things. And I'm beginning to come around to the idea that that is like a reality that. Is gonna happen. Um,
Brendan:Yeah, I'm feeling that too.
Ryan:I'm now like, all right, so what does that mean? You know, for my business, it's like, all right, well, we're, we need to start adopting things now, I I I, I may not be able to see the full picture yet, but I need to experiment and I need to understand what works and what doesn't.'cause if I wait, then I'm gonna be behind. And I'm sure a lot of people are feeling that way, but I, I, you know, on my mobile phone again last night, I was like, all right, what is the next thing I can do? Right? And I was like, all right. I, I, I was thinking like MCP and APIs are gonna be like table stakes in the not too distant future for any company. And you're already starting to see like the, the, the leaders, um, out there on Twitter being like, I, I'm not going, I'm not gonna name names, but someone, uh, that we all know says like, my purchase decisions, if they don't have an API, I'm not buying it. Right?'cause I can't
Brendan:I, I saw that tweet. Yeah, I know what you're talking about.
Ryan:I'm like, you're not wrong, you, you're way ahead of schedule. Right. But I, I think people are gonna get to that point, like even average, you know, non-developers are gonna get to that point where I can't use Claude with your app. Like not gonna happen. Right. And maybe it's GPT or whatever. Right. That, and insert whatever brand you want, but. So I'm, you know, I opened a PR last night. It's like, all right, add, you know, a, a fun function calling tool chains to, you know, for these five things in my app, and integrate that with my internal chat, it in a way that can be extended to MCP when I'm ready.
Brendan:Yeah.
Ryan:And surprisingly, that happened, you know, that took about five minutes, which is like, well, me. Um, but I was like, all right. I. This has legs, right? That wasn't so hard. Like what other use cases are there? you know, so I don't know if there's a question there. Just more of a sharing like, there, there
Brendan:Yeah.
Ryan:I, I'm, I'm now past. What does my development workflow look like? And I'm into, you know, both how can I use it at my company internally and how can I use it in my product externally, and how do I get there in small chunks so I don't, uh, try to boil the ocean all at once. But, and moving forward, um, I dunno if you have any thoughts about that, if you have similar thoughts or don't care.
Brendan:Yeah. No, it's all really interesting stuff that I'm. Also thinking a lot about, I mean, couple things before we move on from what you were just talking about, I don't know if you saw yet, I don't know if it was today or late yesterday, but Google and their workspace, CLI, which I don't know, I don't know enough to, to have an opinion if it's a complete replacement for MCP or whatever, but it's, it's essentially. A way to be more efficient with your tokens when interfacing with Google Workspaces. So that alone, you know, has the potential to blow away tons of SaaS that are out there. And not yours specifically, but just in,
Ryan:maybe, right? Like I still feel like there is value in how the pieces fit together,
Brendan:right. Yeah, I mean, I, this is me just speculating that, like, just the ability to do that. Could potentially be a huge game changer alone. But to, to your point though, like I think everyone does need to start thinking along the lines of whether it's a CLI thing now, or MCPs or APIs, like you have to have a way to interface. So I'm on board with that a hundred percent. Um. But I don't know what the future is either. I have no idea. It's, it's getting to the point where I, I agree. I don't think jobs are going away for developers at least, but I think the industries, industries can, plural, lots of industries, if not all industries are going to completely change in ways that no one knows yet. Um. I, I am a hundred percent on board. Like we gotta be doing stuff, whether we're doing the right things or not. I think at least we're doing something is moving forward. Um, you can't just not, I think that's the wrong place to be right now. We could talk about some of the other stuff that I've been doing that is non dev related.'cause it's all AI related and maybe this is the right time to just talk about it because. The other things that I'm doing all have to do with how do I make myself more efficient so I can keep up and maybe do some of these coding agents myself and teach my team, like we talked about with John and the things he's pushing on, but not completely lose track of all the other things I have to be doing. You know, I'm leaning on cowork a lot because it has the ability to integrate with Slack. Notion, I mean, talk about the brain thing. It, it kind of can do some of that. It's not what you were just talking about where you'd interface with your clients, but it's like an internal brain where I can pick, you know, in our company we've got a lot of stuff in Google Docs, a lot of stuff in Notion, of course we've got Slack. It can essentially, I can ask it a question about our company and our tools. It can find that information in there. Now I'm on the teams account of Claude, so that might be like a, a teams thing. We just added that this week. Um, I can't speak to it too much because I, it's something new to me, but that's one thing I, I'd seen. There's the scheduled jobs, which I'm using a ton of. Now I have my, I have a scheduled job to go out and check Rollbar, which is kinda what I talked about earlier. It goes and checks rollbar daily. It goes and checks my email daily. It goes and checks Slack messages and even like specific Slack channels, and it just gives me summaries, so I'm not missing things. I'm already looking at most of these messages anyways. Although email I'm, I'm not great about for whatever reason. It's, it's been a struggle for me because so much of the other work is done in Slack. But, you know, just those sorts of things that make me more efficient in other areas where I don't have to spend all that time. Necessarily going through all those things, or rollbar or whatever on top of the improvements in development.
Ryan:your work like? I mean, do you like it? What do you have fears about it? Having too much access, like where do you stand?
Brendan:I am not too worried about the access stuff, uh, because I think most of what it's doing is just, you know, distilling things and spitting it back out at me and, you know, at least based on what the team's plan, uh, and I think it's even the, the regular cloud max plans and all that, what their privacy security terms are like. They're not taking that data and they're not feeding it into their algorithms and, you know, training data and all that stuff. Um.
Ryan:was more, I, I've seen at least a handful of horror stories where it's like, organize my family photo. And that translated to deleting the entire, you know, folder, right? Like,
Brendan:Well, for one, I'm not, I'm not going that route yet.
Ryan:yeah.
Brendan:not touching any personal stuff. Um, it's literally,
Ryan:as an example. I mean, like, family photos would be devastating from an
Brendan:yep.
Ryan:Right? But like, deleting, you know, my business, you know, folder would be worse. Right? Like,
Brendan:Right, right. Well,
Ryan:concerning.
Brendan:well, the way, yeah, I mean, the way that I'm using it, it's, it's mostly just analyzing information and. It's not what it does, what Cowork does for me anyways, is I will have it create folders for that sort of stuff and it'll create new files. But I'm, at this point, I'm not going out there and saying, you know, go look in this folder. Go look at, you know, like, I sure Google Drive, but I'm not saying do anything with the stuff besides look for it. Essentially it's like a, a search or, you know, summarize the changes in the last day.
Ryan:Yeah.
Brendan:Yeah. And this is like the tip of the iceberg probably. But you know, I'm also not doing open claw, which has
Ryan:Right,
Brendan:more access and letting someone, someone else's open claw or even a person come along and say, Hey, I'm, I'm the FBI, gimme all your information. And then it just hands it over, right? Like, I'm not anywhere near that level.
Ryan:Yeah. Maybe that's what I was thinking of. Open claw versus cowork. But
Brendan:Probably 'cause 1, 1, 1 thing I'll just mention, like when I, I was trying to think what I did. Oh, that's what it was. I, I tried to have cowork do something in Gmail. Um, gosh, I'm, it's the tip of my tongue. It wasn't anything destructive, but it was also just like, oh, that's what it was. I was like, can you just archive all of the, the things you just summarized for me so they're not in my inbox? And its response was, no, I can't do that. Where OpenClaw probably would have full access to that sort of stuff. So that maybe that was the difference you were thinking about.
Ryan:Yeah, definitely. The stories were about open claw, not cowork now that I think about it. But
Brendan:Yep.
Ryan:yeah, I mean something, I, I haven't played it along with cowork at all, but all that sounds interesting, as does like this notion, you know, we, one of, I think it was our second or third recording we talked about, um, the, uh. The para method and how to organize your second brain and, you know, different tools like obsidian. And I feel like those things are coming back in favor now, that you could be like, here is my second brain, but I don't have to like, worry about the, where the information is stored. Just give it, give it to Claude or whatever model and be like, this is what I'm looking for.
Brendan:Yeah.
Ryan:Give me the info. Right. It it's be is, that is interesting to me on a separate note.'cause I always, you know, I never keep up with the, with the second brain stuff,
Brendan:Yeah.
Ryan:of resources. I'd love to figure out how to do that. And maybe cowork is, is part of that solution.
Brendan:Well, I mean, just to touch on that a little bit, you know, you put me onto the para method after that episode and I've tried and we'll say partially successfully to have like the structure that that second brain concept. You know what, what he talks about in the book. Um, so for instance, when I create those folders I was talking about, I'm essentially putting them into like my projects or my resources or my areas folders that I already have.
Ryan:Yep.
Brendan:And so it's sort of following that structure, but I think what you're getting at is not even necessarily needing that, but allowing this to basically operate that way. There is a productivity plugin, and that's the other thing too, like there's tons of plugins that can, you know, help it. Be a project manager, whatever, all these different things, but productivity is one of 'em. And that kind of does start building its own little brain where it's, it's just MD files, you know, it's, it's the same kind of thing that we've been doing for a long time. Not, well, I guess not even a whole year with Claude Code yet, but same concept where it's like, here's the file to reference. So for instance, like. The initialization there was basically like, who do you work with? What are their names? And this stuff's all stored locally, so it's just referencing those files. It's not like putting it up into the cloud, but, you know, and now if I, I think it can even do summaries and all these different things where it'll keep layering on into there. And then it becomes my place where I could also just say, Hey, what's going on? What, what was that thing I talked about last week around X, Y, and Z or whatever. Um, so yeah, they're, they're doing some of that stuff already.
Ryan:yeah, I've started to get back into rag and, and function calling and tooling
Brendan:Yeah.
Ryan:and starting to try to really understand the difference. But I feel like, you know, you, we, we have this wealth of unstructured data where RAG, I feel like is like the a, a pretty good solution to pulling context into
Brendan:Yeah.
Ryan:that sort of thing. Um. I'm definitely playing around with that a little bit more, uh, these days. So we'll see you. there's a lot of fun stuff happening. I had like an extra day every week to just work on that stuff.
Brendan:Same here.
Ryan:Um, but in the meantime, I'm trying to do a little bit of ATI a little bit at a time and just continue to explore.
Brendan:Yep. Yeah, same here. Um, I don't really have much else to add. I guess that's, it's a lot of stuff. It's been kind of a whirlwind even the last week of all that stuff that we've been doing and I've been doing, so
Ryan:Yeah. Well,
Brendan:I.
Ryan:say, you know, if, if you're listening to this and you have thoughts on, you know, what we've been talking about or how you're approaching things, you know, feel free to reach out. We'd love to chat with you about this topic and see what other people are doing. Uh, 'cause I think a lot of people are interested in trying to figure it out together.
Brendan:Yeah, and I'll even follow up with that, that like if you don't wanna be a guest or you have any reason not to, like, feel free to just email in and we can read those emails potentially and talk about 'em as well. So just wanted to throw that out. There's another option.
Ryan:That's a good idea. All right,
Brendan:Cool.
Ryan:it there.
Brendan:All right.
Ryan:again, uh, soon and we'll, we'll touch back on the conductor stuff and see where we end up.
Brendan:All right. Sounds good. See.
People on this episode
Podcasts we love
Check out these other fine podcasts recommended by us, not an algorithm.
Code with Jason
Jason Swett
Standing in the Fire
Very Good Software
Remote Ruby
Chris Oliver, Andrew Mason, David Hill
IndieRails
Jess Brown & Jeremy Smith
Startups For the Rest of Us
Rob Walling
Vibing with AI Code
Brendan Buckingham