Episode 509
509: How AI Fleets Fixed 31 Issues in Two Days
April 6th, 2026
42 mins 11 secs
About this Episode
In this episode Frank and James riff on everything from accidental source‑map leaks (yes, even Apple slips up) to using Copilot and multi‑agent AI workflows to triage issues, write PRs and even plan trips. They share hands‑on tips—use “plan mode” and deep research, turn plans into issue artifacts, run multi‑model code reviews, cap concurrent agents (~3), and babysit agents—to boost productivity while avoiding chaos and burnout.
Follow Us
- Frank: Twitter, Blog, GitHub
- James: Twitter, Blog, GitHub
- Merge Conflict: Twitter, Facebook, Website, Chat on Discord
- Music : Amethyst Seer - Citrine by Adventureface
⭐⭐ Review Us ⭐⭐
Machine transcription available on http://mergeconflict.fm
Support Merge ConflictTranscript
James 00:00
Welcome back everyone to Merge Conflict, your failure developer podcast in the entire world where we just ship all the bits and just merge and ship stuff to NPM accidentally. Whoopsies, what happened? Just ship all the maps. We can find everything with a map.
Frank 00:13
Oh. I've never heard of map files. Tell me more, James. Did someone accidentally upload a map and now everyone's reading a map? That sounds ludicrous. Why would anyone want to do that?
James 00:25
I did watch a video from, I'm pretty sure it was the the guys at Syntax. They were talking about source mappings and why you may want to because I think Apple. Remember when Apple, they shipped their source map accidentally for the new Apple App Store page where you can browse apps on the website.
Read Full Transcript Hide Full Transcript
Frank 00:40
Oh.
James 00:45
And they accidentally shipped their source maps.
Frank 00:45
Yeah. Yeah.
James 00:50
To me, these are like PDB files in a way, but like different. I think they're like also like the entirety of the source code kind of, but not really. The source maps are important because they kind of allow you to like really do deep debugging from my understanding.
Frank 00:58
yeah
James 01:03
because it's like the source and it is the map X, Y, Z to it. It's not like a one-to-one writer. Is it a one-to-one? I don't know.
Frank 01:10
No, think of it more like as the de-obfuscator. So like you take your JavaScript file, you obfuscate it, and it's the information you need to turn it back into source code. So everyone ships their JavaScript apps as obfuscated, compressed, compiled, whatever silly web term you want to use there. And it you know it it turns all the variable names into underscore A or something stupid like that. And then the map says underscore A is actually the variable called prompt, let's say.
James 01:45
Eww.
Frank 01:47
Yeah, they're they're useful um because, yeah, it's you you want to test then actual release bundle because bundling JavaScript is such a pain in the butt and getting that all right is a pain. So you actually kind of want to test your app in that compiled mode. Plus there are efficiency gains, but they're very minor. um But you want to be able to read the code in the debugger. And that's why I think you think of it as debuggable because yeah, it puts the variable names back, puts the white space back, all the important things.
James 02:19
Yeah, it's good to do that. And I think obviously if you're doing like internal testing and obviously running on your mobile machine, that's super important, but you probably don't want to ship them to production, although you can. i mean gen i think, could.
Frank 02:30
Yeah, I mean, If you don't care, like who cares? Like software's free now. I believe I heard someone at Microsoft say, so like what's the value in anything anymore? You might as well just ship all the code. Now I got i think like is if it's open source code or if it's just like a bunch of front end stuff that's, you know, validating a text box, who really cares? You don't need to obfuscate that stuff. But if it's all your API keys and core logic to your application, maybe, maybe you want to protect it.
James 03:03
If it is cloud code, maybe you want to protect that accidentally. And it's an easy mistake to make. But yeah, there's a whole thing. I didn't dive through any of the code. I don't even know. I think all the code's taken down or something. Someone's converting it to Rust or something crazy.
Frank 03:16
and
James 03:17
But yeah, think in world of... in the world of software development, right? it's It's a great reminder. Apple did and Anthropoc is some of the largest companies in the entire world. It's easy to to to slip up a little bit here and there and that happens, you know, but also give people grace. Like there's a human, there's a human that that accidentally flipped a flag, right? And that kind of sucks for them, but it's a bad day for them.
Frank 03:41
chance it was a human. 90% chance it was a human. But it could have been anything accidentally flipping that flag.
James 03:45
ah or a human instructing, human in the loop.
Frank 03:49
Right.
James 03:49
There's gotta be a human somewhere. there's a human somewhere, Frank, that's human.
Frank 03:53
Someone has credentials. We'll put it that way.
James 03:55
Yeah.
Frank 03:56
Whoever's credentials is used. Yeah, you know, it's funny because i was just going through a whole thing.
James 03:58
But talk,
Frank 04:02
um was getting the weirdest.NET compilation error. like get it when Whenever I was trying to do things in CICD, it kept hitting this weird firewall rule. And it would keep coming up in my CICD. Everything was working. But this firewall is like, we blocked your access to this domain I've never heard of. I'm like, what the heck is that? And finally tracked it down.
James 04:24
Weird.
Frank 04:25
i was using it. I was even using it. There was an ancient, ancient obfuscator referenced in in the repo. It wasn't even referenced. It wasn't used. Nothing was executing it, but it was sitting in there there in the repo because i am dirty. I am sloppy, and I just leave stuff in the repo from decades ago. And um it was AIs futzing around promiscuously trying to load yeah URLs. was trying to load an obfuscator yeah URL that is just so ancient that like it was hitting all the rules and everything. I was like, oh, yeah, I remember when I used to obfuscate my C-sharp code.
And I think people still do that, but I don't know. No one decompiles code anymore, or at least you know at given the web and how every app is basically open source because of JavaScript. What's the difference?
James 05:20
Yeah, it's tricky. I would say that that is, you have like an idea and people then you can just, you build it. It's easier to build it faster than ever nowadays. If you have an idea or you see something, um I think back in, well I think in the enterprise it's pretty important because they really want that extra layers of security, right?
Frank 05:39
Yeah, I guess.
James 05:41
So especially on like Android and like obviously with,
Frank 05:43
Or at least the theater of it.
James 05:45
Yeah, there's a theater. there's a There's like a feel-good feeling that i've I've obfuscated my code and I went through all this stuff.
Frank 05:53
Yeah.
James 05:53
And look at all these look at all these checkboxes I checked, boss.
Frank 05:54
Uh-huh.
James 05:57
No,
Frank 05:58
There's absolutely no way an AI could just sit there and give it good variable names. That would be impossible.
James 06:05
impossible-able. Well, and that's, you know, we talked about when you disassembled that thing that fell off the submarine or whatever, the boat.
Frank 06:10
Mm-hmm. Yeah.
James 06:12
Was it a boat? Was the submarine? What was it?
Frank 06:15
A frigate.
James 06:16
A frigate? It was a frigate. You know, you nowadays you can point it at the thing and then figure out the thing. So I think that's a beautiful world that we're living. I was, we're at MVP summit in, you know, aerial tramways are having, having a moment, thanks to SNL.
Frank 06:28
Oh yeah, that was great.
James 06:29
and i And didn't know much about aerial tramways. And ah yeah i think the thing is I can point AI and say, give me a thing about aerial tramways.
Frank 06:33
Funuculars.
James 06:38
And then I can have a generator report, a PowerPoint.
Frank 06:38
Hilarious.
James 06:40
We um got invited to this wedding ah that's um um in another country, in a beautiful country. I'm not going country, because this thing's a beautiful country in South America.
Frank 06:49
Cool. There's many to choose from. Yeah.
James 06:53
South America. Latin America, South Latin America.
Frank 06:54
Ooh.
James 06:55
Latin America. And um I was like, well, I want to go to this wedding center, been in this country. So I went into Microsoft Copilot, normal Copilot, and they have this new tasks.
Frank 07:03
oh nice oh
James 07:10
And our friend Craig works on the MAI Copilot team. And you can go into tasks and you can ask it to do a bunch of stuff. So for example, I had a book a date night for me and Heather. And opened OpenTable, it did a thing, found a restaurant, did a thing, right?
Frank 07:23
oh
James 07:25
I'm going to have it book super, mm-hmm.
Frank 07:25
Wait, wait. You did the actual demo thing where you had it book you a dinner reservation? mean, that is the classic AI demo. I'm so proud of you. I've never done that.
James 07:34
Yeah.
Frank 07:35
ah
James 07:35
I was like, Hey, find some restaurants that we haven't been to and go book it for me and make it happen.
Frank 07:39
Yeah.
James 07:40
I said, I said, you know, um, you know, you know, want to have dinner around, you know, but six o'clock.
Frank 07:40
Cool.
James 07:45
So we got to be home by like so eight o'clock for Millie for insulin and, you know, you know, see what nights are good. I prefer a Friday, but a Saturday would work in these day range. Like I'll make it happen.
Frank 07:55
Mm-hmm.
James 07:55
And it, it made it happen. and i was like, cool. Thanks. Awesome. I just did it.
Frank 07:59
I love it.
James 08:00
Did the thing.
Frank 08:00
I mean, the next thing I know, you're going to order a pizza and book a flight. like it's I mean, you're using AI so correctly at this point.
James 08:06
It's close.
Frank 08:08
Yeah.
James 08:09
So for this trip, I said, Hey, we got invited to this wedding.
Frank 08:09
okay.
James 08:14
Here's the wedding website. Here's where we, you know, you know, they know where I live, obviously, because I've told it, it's got the memories.
Frank 08:15
Okay.
James 08:19
And I'm like, Hey, you know, Heather and I need to go, like, go figure out optimal flight. I want the best flights, reduced hours.
Frank 08:27
okay
James 08:27
I'm okay with an overnight, but like, we want to spend some time in the main city before going over here.
Frank 08:29
Jeez.
James 08:31
I said, put together an itinerary of everything we need to know for the wedding, getting a flight options. said, I want Alaska, but I also want United because i got I got credits on United. But I'm also cool splitting one way each, right? I said, I got about $500 in credits. you know What can we do?
Frank 08:46
Smart.
James 08:46
um And then I was like, you know when we stay in the city, you know um and in the first city, i want to stay at a Marriott because I got points. And then we'll stay at the the wedding hotel here. I said, go. i said, you know what? it i want you to not just research this. I want you to create a a PowerPoint presentation to convince Heather that we should go to this. And it generated the most beautiful, stunning experience thing in the entire, I'll send it.
It's like, it mind bogglingly amazing.
Frank 09:12
Oh, my God.
James 09:14
was like beautiful imagery and like all this stuff. And then, um, I didn't have it book the things for me. That was like the next thing I could have.
Frank 09:20
OK.
James 09:22
Uh, cause I sat, we don't, we sat down, but it was really nice cause it did give a review. So I've been doing that. So thanks. Shout out to Craig and the entire MAI team for, for hooking that up. was pretty cool. And, um, it was nice.
Frank 09:31
Well. That's very cool, but you just gave me the idea. Now you need to code um an MCP server that logs into all your reward accounts. So it can be like, hey, you got points over here. You got points over there. going to distribute your points to the economy and hook you up with the best date night ever. Mm-hmm.
James 09:51
I do feel like that's like the next level jump of, of whatever this is, is that if it can have access and just even just understand it doesn't a be to have access, but understanding to my points and my things, Like ideally what I want is an agent. I want to be like Hey, I want to go to, we just went to Korea, Korea. You know, we want to find fly Hawaiian airline and we got this Alaska airline. We got this many points.
Like we want to go sometime for this long, blah, blah, blah. Right. I want you to like monitor, and optimize the absolute absolute best you know point to dollar ratio spend upgraded business class blah blah blah i don't want to spend more than this point and then it could just work in the and it's just an army things or or it it could just ideally it should just know and it could just give me recommendations and non-stop right there's services that i have like there's um
Frank 10:31
Yeah. Love it. Right. and Yeah.
James 10:48
what is it a points path, which when you search on Google flights, it'll tell you like the points of like if it's a good deal or not. But you got to be proactive. I listen to all the hacks and he's always like, oh, here's how I do this. I'm spending lots of, years I don't have time to research, right?
Frank 11:01
Yeah.
James 11:02
So our flight back, we had this like, you know, point, I was like, oh, it seems like a decent amount of points.
Frank 11:02
Yeah.
James 11:06
I have no idea if it's a good idea, but i was like, seems all right. It doesn't seem like too much. Well, I'll get it.
Frank 11:10
Mm-hmm.
James 11:10
yeah Let's book it, right? But is it a good dollar to point ratio? I have no idea, but I do think that that's the that's the the agent that needs to sit there and whoever builds that billion dollar idea, I'd pay for that.
Frank 11:16
since I've got half a feeling like Open Claw does it because, you know, there there's skills for everything.
James 11:24
Probably.
Frank 11:25
But yeah, someone get on that, please. And I'm just realizing, like, I actually do need to start taking advantage of this because living on the island here, all I think about are logistics when I leave the island, like timing this and timing that. got to time the fairies, got to time this. It's horrendous. I hate logistics. I'm willing to do it for very important events, like to go visit you, but that's about it. And so I'm like, oh, but if an AI actually did that and maybe coupled in my points, which I don't have any of, but you know I could pretend we could start building some points together. uh yeah i do need to get it on logistics thanks for the uh tip i always think of it as ordering pizza but now that you've reframed it in my head as getting on and off the island i think i might actually try to tap into this
James 12:12
I did think about this recently and like recently I did a door dash. I don't door dash very often, but like I was, I was, i was remote. I was in a place where there was no food in the house. that we were at temporarily. And i was like, I gotta order some stuff or some bars and some, you know, some waters and some stuff from seven 11, got it delivered. You know, I didn't have, but I couldn't leave the house.
Frank 12:34
Classic.
James 12:35
And this weekend where we're working on this, in this, in this area. And I was like, it's like three o'clock. We haven't eaten anything yet or anything in the house. I was like, I'm just going to like reorder. But I like, because I don't like apps, like I'd already deleted, I'd already deleted the DoorDash app and I got to reinstall, I got to like re-log in, I got to go pick it up and do it, want everything.
Frank 12:48
OK. Hit me again. Yeah.
James 12:55
What would have been cool is I, and I should have done it, go to task and be like, hey, go place that last order from, it's the same, but this address, like go, make it happen, right?
Frank 13:04
Yeah, I love it.
James 13:04
And I bet I would have, it might've been able to do it. would have been good test, you know, how close does it get? It gets pretty close until I also like enter a credit card, you know, so.
Frank 13:13
Yeah, exactly. Well, we both keep talking about, i think at the MVP summit, you called me skillless. And it's almost true. I have one skill. But it sounds like we both need to build up our little skills so that we can do this kind of stuff.
James 13:28
Well, it sounds like somebody to deployed ah a agent fleet onto a code base over here.
Frank 13:29
We'll just call it.
James 13:36
I don't know. You came back from MVP Summit and all of a sudden Frank's using all of his tokens. And that's the 31st. guess you guys it's about to reset. um So C language, Frank's C++ plus plus parser, compiler, and interpreter for.NET, which think you use this in iCircuit, right?
Frank 13:54
iCircuit, yeah. So it it simulates Arduinos in iCircuit. So you can just write Arduino code in iCircuit, and while it's simulating the circuit, it's also simulating the Arduino code. And so I have a full high-performance virtual machine written in C Sharp that runs on.NET on any platform, all sandboxed up with a little C compiler. So you know it's an open source library because I think like I don't know who else could use this, but it's good for parsing header files.
If you wanted a plugin system for your own apps and you were willing to have people write those plugins and see in C and C++, which is a little weird, um it's great for that.
James 14:34
Well, that doesn't matter anymore.
Frank 14:35
But...
James 14:37
Anyone can do it and just
Frank 14:38
Exactly. Right.
James 14:39
Yeah, just but do it.
Frank 14:39
Just have an AI generate the code.
James 14:40
Yeah.
Frank 14:40
You might as well.
James 14:41
yeah
Frank 14:42
Then there's a plugin system. I make it super easy to like integrate the the the sandbox virtual machine with your own code so it can talk to like your own objects and everything real easy. But yeah, I've had this, ah it's ah some old artisanal code I wrote by hand back in the day. Yeah.
James 15:00
Last last last ah release before today was almost two and a half years ago, by the way.
Frank 15:03
m
James 15:06
And original releases are 2018.
Frank 15:07
I mean, yeah, it's like, well, the C language hasn't really, OK, the C language changes constantly.
James 15:08
So this is super artisanal. eighteen so this is super artesal
Frank 15:19
But like the important parts of the C language don't change that often, so once you have something working. So the deal is this library was a C parser compiler in virtual machine. But a lot of Arduino code out there is c plus plus And C++ plus plus is a much bigger language than C.
James 15:37
Ah.
Frank 15:41
Much harder to implement. C is this beautiful, tight, easy to implement language. C++ plus plus is the monster from the middle of the ocean and coming to conquer your town with its nuclear ray breath and all that kind of stuff. That kind of monster. Kaiju. That's the word I was looking for. C++ plus plus is the kaiju. Anyway. ah So I was basically goofing off all last week at MVP Summit. And I came back and I'm like, I have not used up my tokens. And I swore i was going to use all my tokens every month. And so I'm like, what am I going to use all my tokens on?
James 16:17
Yeah.
Frank 16:20
So I did two things. um We talked about it a little bit at the and MVP Summit episode, which everyone you should watch live because we were facing each other. It's always very awkward and cool. um I actually implemented my agentic workflow, which is really cool. It goes through past issues and triages them. That's that's a side point. That was not using my tokens fast enough, James. I had tokens to burn. So I'm like, you know what? It time. is time to improve my C language and turn it into language. Or at least get in some of the bigger features it's been missing for a while, namely like inheritance and polymorphism, operator overloading, the stuff you need to kind of support the more sophisticated Arduino libraries out there. Because no one writes in C. Everyone writes in C++.
Um, so you say I ran a fleet. Call me Admiral Frank because I did not use the fleet command. I was one at a time banging through these puppies, one PR at a time. Also, ah Pierce made me feel bad at MVP summit because he was asking who, who launches an agent and then sits there and watches it instead of multitasking. And like, that's, that's me.
James 17:38
Everybody.
Frank 17:38
Hi, hi. Yeah, because, look, AIs have improved. but They really have in the last year. I get it. Opus, awesome. But they still make mistakes. They still take every little shortcut they can possibly take. So, like, you do have to babysit these things a little bit. Anyway, Pierce put me on the spot, and I said, I am going to... I'm going to multi-agent James. I was creating issues. I had agents creating issues. i was creating PRs. I had agents critiquing those PRs. was agentifying everything.
And I... I, you know, not to over ah compliment myself here, but i crushed it, made some awesome features and really improved the library and full credit to Opus for doing all the hard work, obviously. But I feel like that was a blitz of dev lead management. um I'm kind of exhausted after just like two days, but the agent and I, we accomplished a lot. And it was honestly really thrilling and fun because like it just got so many big features done and bugs fixed that have been bothering me for literally years. like Some of the issues were five, six years old that we fixed. Anyway, that's a long rant. I just want to say I had so much fun. like This has really inspired me again to like, you're not coding, but it's coding. It's weird. It's a different kind of coding.
James 19:01
It's coding. is is is there we I think this is a cool example because you know now I can see in your repo you got an agents MD file, you got some like workflow files, you know you're trying to agentify this puppy.
Frank 19:02
It's coding. Yeah.
James 19:12
You got the releases and like it touched all aspects of it. and Funnily enough, like One of my new favorite things to do is what you just did here, which is like go find one of my projects that I just haven't touched in a long time and can I facelift it, right? It's still deployed, I'm still using it, maybe it's an app, it's whatever.
Frank 19:30
yeah
James 19:31
um Maybe it's an open source project. ah Recently, there's a project on the my on the Microsoft GitHub called Cloud Cloak, okay?
Frank 19:39
Cloud Cloak.
James 19:41
Cloud Cloak, it's github.com slash Microsoft slash Cloud Cloak.
Frank 19:41
It's hard to say. Cloud Cloak.
James 19:47
What Cloud Cloak lets you do, is's in a it's a browser extension. And it is when you log into GitHub, into AWS, into Azure, into other services, it will blur all of your API keys, subscription keys, all that stuff for you automatically, right?
Frank 20:04
Oh, fancy.
James 20:07
This is a hackathon project a few years ago, pre-AI.
Frank 20:07
Mm-hmm.
James 20:10
maybe maybe early co-pilot days of like auto-completion.
Frank 20:11
Nice.
James 20:15
And a bunch of people worked on it.
Frank 20:15
Mm-hmm.
James 20:17
Scott was like a sponsor on it, like real excited about it and it's great for demoing because it can auto-hide a bunch of stuff even on your screen in the browser.
Frank 20:22
Right. This is in the browser, I'm assuming? Yeah. Mm-hmm.
James 20:28
So you just turn it on. It'll auto-detect. has a bunch of yeah URLs that'll auto-detect and turn on automatically. was like Azure, AWS, or things like that. And it's like, oh, flip it on. you know, if you decide to enable it or not.
Frank 20:37
Yeah.
James 20:39
um And this is super great. So for example, even if you have an API key, you can say show API key, it'll still blur it, but the copy button will be available for you. Right. So that's kind of nice.
Frank 20:49
Yeah. Great. Yeah.
James 20:50
Um, uh, so this project was sitting there and I got a report, uh, from our video editor. It's like, Hey, like, this is a pain in the butt to like edit it out and blur all these things. Like there's gotta be better ways.
Frank 21:04
Yeah.
James 21:04
Like, Oh, there's this cloud cloak. And I went and I looked at the repo and I was two and a half years old and like no one's touched it.
Frank 21:10
Yeah. Well,
James 21:12
And I was like, can I get the admin?
Frank 21:14
ah if it works.
James 21:14
Give me the admin. Give me the admin.
Frank 21:15
Yeah.
James 21:15
Give me the admin privileges.
Frank 21:16
Okay.
James 21:17
And I unleashed.
Frank 21:18
Mm-hmm.
James 21:18
So I pointed Copilot and I was like, I said, I want you to analyze. And there was about like 15 open issues. And I said, I want you to, and there was no release pipeline, no anything. So I was like, first and foremost, like there's no tests. I was like, let's get a baseline test. Boom.
Frank 21:31
the
James 21:33
And was like, all right, I want you to analyze all of the issues. I want you to understand what's still valid, what's not valid. And some of these don't have a lot of information. Can you research and like based on the code base, update the issue to see if there is more information so we could triage it. And I said, I want you to put together a plan, a phase attack plan of like, what's this boom, boom, boom.
Frank 21:49
Love
James 21:54
And it just got to work. And like within one day, i was just like, I i did so much. fixed up. I added tons of new Azure websites. I like fixed so many bugs on it. Just like you're saying, it's just like this army of, I've never i've never seen this code. I've never seen this code. Right? And it's just like, boom, it's like doing the things.
Frank 22:11
love it.
James 22:13
i was like PR in it and doing all this stuff. I was like, dude, crushing it. It was fun. It was fun. it's It's coding a different way.
Frank 22:19
Yeah. Yeah. thanks Thanks for making me feel better because like i I did, I did have to do a tiny bit of code patch it here or there, but like for the most part, this was all AI code. Like if it did a bad job, I just yelled at it. um I've really gotten used to, um this is another reason to babysit the agents steering. um I didn't know that this feature was added.
James 22:39
o
Frank 22:41
it It was added to VS Code, but it's also sitting there in Copilot on the GitHub website. So when you when it starts a PR, you can click view session and you can monitor it, see what the heck it's doing. But there's also a little chat window in there. And if you see it doing something really dumb, you can just type to it and say, hey, buddy, stop messing around with that. Do it this way instead. And hit the button. And they say on the next tool call, they'll just kind of inject your prompt in there. So you're not like stopping it. You're not restarting. And it's you're just kind of sliding a little message, and sliding into the DMs of the AI there. And so I found that that was a really great way to steer it when I saw it going wrong.
And then obviously you go patch your agent's MD file. So it doesn't make those kinds of mistakes again. But boy, it was gratifying, especially like the way it was building unit tests and everything. And there were features like in the library that just like, I know I implemented this feature, but I don't know if it's complete or not. Can you just write a test suite for me? yeah.
James 23:46
Yeah.
Frank 23:47
like check if it's finished. Cause like that, that was one thing I learned when I started my agentic workflow, which has been really beneficial where I ask it, um go through the old issues and just do deep research. on don't know why i keep using that term. i I think I heard it at um MVP summit.
James 24:04
a good, yeah.
Frank 24:05
I just liked it.
James 24:05
Deep, deep research.
Frank 24:06
I'm like, like do deep research on this.
James 24:07
I mean, you can do... you can do You can do research, but then you can deep research.
Frank 24:12
Right.
James 24:13
like the Research, that's garbage.
Frank 24:13
That's what I'm saying. Yeah.
James 24:16
That's 2025.
Frank 24:16
yeah
James 24:18
We're We've got do deep research.
Frank 24:18
Right.
James 24:19
Deep research.
Frank 24:19
ye I need deep research here.
James 24:20
Hmm.
Frank 24:22
And I want you to come up with a test suite. I want you to do this and that. And it's really satisfying to see, like, generate those tests. Because I'll be honest, James, like, um we used to say at Microsoft, devs make the worst testers. And I do believe it because we're so happy when we get our own code working that we don't really want to break it. We say we want to break it. We don't really want to break it. don't want to break my code.
I know my C compiler has issues and I don't want to look at those issues because they're big and scary and compilers are hard on us.
James 24:46
Yeah.
Frank 24:52
But the AI is just like, no, um ah ah we're going to do multi-base class inheritance. And i'm like, no, that's not going to work. It's like, yeah, it's probably not going to work, but we're going to test it. I'm like, no, don't test it. It's not going to work. and it's like, oh, it doesn't work. Let's implement I'm like, how about you implement it? And then, you know, life is good then. Yeah. It's actually, i'm I'm lying about that one.
It really didn't want to do multi. and For those who don't know, C++ plus plus has multiple base class inheritance. You can inherit from multiple different types all at once. And basically every other language has dropped that kind of support. But C++ plus plus has got it. for better for worse and if you have a c++ plus plus compiler you got to support it yeah so it's funny like even the ai is like let's not do that i'm like let's do it let's do it i got i got tokens to spend baby and you are you have no feelings as far as i can tell i anthropomorphize you but i i don't believe you have feelings so you're going implement multiple base classes and it was great it's it's just um what did i say i think we got like
James 25:32
Yeah, it's bananas. and That's crazy.
Frank 26:02
31 different issues resolved and fixed in two days.
James 26:04
It's a good day. Two days? Yeah.
Frank 26:08
Two days, two days. And i have to say, like tiring days because multitasking is hard is what I've learned. like you've You've always got like one running, one that you're reviewing, one that you're thinking about running. And I find about three is about as much as I can mentally handle. I see that like tweets of people with like a grid of 27 processes running. don't know how they keep all those in their head.
Or how do you verify any of them? because like The AI kept making mistakes. I had to keep telling it, like, no, no, stop cheating.
James 26:38
Yeah.
Frank 26:40
um
James 26:41
Yeah.
Frank 26:41
But like I found three is a good number.
James 26:45
I think like that's a good number, especially if you have like certain things in your project that you know. i was talking to someone like, well, how do you do it? like How do you like have all these sessions? was like, well, I kind of understand the code. So I can look in triage like in my mind, either features or issues or bugs.
Frank 26:59
Mm-hmm.
James 27:03
And I can run things in parallel, even on the same branch. If I want to if I'm just working at main or on multiple branches, I find work trees and I want to avoid merge conflicts.
Frank 27:07
m OK.
James 27:13
So in my mind, like you can have the AI prioritized, but in my mind, if I know the code, which is beneficial, and can then have the ai work on different pieces of the code that aren't going to conflict with each other.
Frank 27:14
Yeah.
James 27:23
Right.
Frank 27:24
Right.
James 27:25
Hopefully. Right. obviously it can, it can resolve merge conflicts and be all smart about it.
Frank 27:27
Obviously.
James 27:29
but then that gets another cognitive load. That's another thing that I don't want to deal with. Even like if I'm just writing code manually, I don't want to handle merge conflicts personally.
Frank 27:37
Right. Yeah.
James 27:37
But then I also said, you know, I'm not just working on one project. I'm working on other projects too, right? so in an ideal world, I have like, um, you know, multiple VS code windows open or get up.com open or whatever it is, whatever app is open. And I'm working on multiple things at the same time. Not that I'm focused on all of them, but if I have ideas or I have a backlog item or have something, maybe while I'm waiting around, I'm like, Oh, let me just go check on this other project or do this other thing. And I kind of sat around today.
And I was like, I don't really have anything to work on. so i was like, that seems problematic.
Frank 28:11
Oh, wow. Yeah.
James 28:13
you know At least on my current projects that I have. and i said, that seems problematic. i need to like Maybe I need to go tell the AI to come up with new features to implement, right?
Frank 28:17
Yeah.
James 28:21
you know And in some workflow. um But then I was kind of like, I'm kind of bored with all my agents getting to work and doing something. I kind of found this flip side where i was like, I was exhausted when they're all doing too much. i have to review too much. And I'm like, now I'm kind of bored. The agents aren't doing anything. I need them to do something for me.
Frank 28:35
a
James 28:36
I need, I don't have a new idea, but like, well, let's go.
Frank 28:37
yeah
James 28:39
So yeah, it's kind of interesting.
Frank 28:41
Yeah.
James 28:42
Yeah.
Frank 28:43
So even though i said I was getting tired with three running and I got 31 things done in two days, I was also working on other projects at the same time.
James 28:51
Oh, cool.
Frank 28:51
So I was working on iCircuit.
James 28:52
Nice.
Frank 28:54
And iCircuit uses two libraries. I mean, it uses lots of libraries. But the two libraries I was also working on were C language, the C compiler, and CrossGraphics, which is the rendering engine used by iCircuit. And forever, I wrote the most beautiful artisanal metal graphics layer in cross graphics, all in the name of performance. Always just trying to like you know get get those frames per second with using zero CPU was always the goal.
um But there were a few quality issues. So I was always had one or two PRs in iCircuit running. three or four things in C language running all on GitHub because I hate work trees. I think they're gross. ah Now that I've had time to experience them, I think they're terrible. um And then i had like, as a fun side thing, if I was waiting for any of those to finish, I'd spin up cross graphics and keep working on the metal renderer where, you know, it's funny.
it's It's those funny conversations you have with the AI now. I'm like, hey, there's a bit of code. I've always been afraid to write because it's really math heavy. I'm good at math, but it's still intimidating. Let's work on this together, babe. And it's like, let's do this. And it generates the absolute worst code I've ever seen in my life. Completely fails, completely breaks every test. I'm like, well, that was way worse. um Can you review your own code and figure it out? And we went back and forth and like we both decided the math is really hard.
in this case but because of that dialogue um we came up with a whole different way to implement the feature that i wanted and it came out much cleaner and much better and it came out actually really fantastic in the end and that was a third side project while working on two other projects each of those projects having well i circuit i could only handle about one or two on at the same time because they're much more complicated, a lot more to think about. And the C language one was just blasting away. Man, I've been productive for the last two days. I don't know if I can do it every day. Oh, every week.
It's too much.
James 31:02
Well, I would tell people some pro tips that I learned from going to Harold's lab at the MVP summit and conversations afterwards, which was like plan more, my personal opinion and Harold's as well.
Frank 31:14
evening
James 31:16
He's like, if you're not planning, you're doing it wrong. And I'm like, Whoa, shit. I'm not planning enough.
Frank 31:18
Yeah.
James 31:19
oh Oh, crap. I crap. Oh, I swore on the podcast. i was like, I'm not doing it enough. You know, um I do plan a lot. And now some of the agents in the main loop will ask you questions.
Frank 31:26
yeah
James 31:29
And I really like that. I wanted to ask me questions. In fact, sometimes if I'm not in plan mode, I'll say, ask me questions if you're not sure, right? um But i do think plan mode always, 99% of the time, results in better results at the end of the game for implementation. And then the other thing that he said, so so be planning. So I'm interested if you were to go back in time and instead of just if you went into a deep plan mode and and going back and forth, I wonder if that would have resulted in better results first.
Frank 31:56
No, I have my own technique now. So I definitely do plan mode when I'm just like coding in VS Code or something. But I have my alternative to plan mode that I very much enjoy, which is I always have it do first a plan mode where we talk it through.
James 32:10
Deep research.
Frank 32:11
Then we do deep research, James. And we summarize the plan, the deep research, and the discussion as an issue.
James 32:15
Yeah.
Frank 32:21
have that put up on the thing and then I assign the issue to a coding agent.
James 32:22
Hmm.
Frank 32:28
So there's almost three levels involved.
James 32:28
Nice.
Frank 32:30
There's the planning, then the finalization of an issue which generates a real artifact that I can reference and think about and critique and play with and then I code from that.
James 32:36
Yeah.
Frank 32:40
And that has been my favorite process so far, because having that like codified issue up there that is like, here is our plan.
James 32:40
like that.
Frank 32:49
This is what we're going to do. And it's done the deep research. So it's all I know I approve of it.
James 32:53
and And like the context is fresh there. So it only needs to know whatever you it did. It doesn't need all that other crap that you threw away in its context window.
Frank 33:02
Mm hmm.
James 33:03
So again, start fresh. The last thing, which I like this approach too, that you just did. um i like, i love creating issues is the last part is the really, the review part of it. Because I think there's the human review, but now I've been using the GitHub Copilot code review agent much, much more. And additionally, what I've been doing in the CLI is I'll point it at a PR and I'll say, hey, there's this PR that just finished.
I would like for you to review it with multiple models. Like give me a GPT-4 model, give me an Opus model, give me a, you know, whatever model.
Frank 33:35
Mm-hmm.
James 33:36
right Maybe I have my own model, my Gemini model. I said, give me these. I want you to review. and What you can do is say, review with these different, because who knows who implemented it? right maybe it's Maybe it's one or the other. right maybe it's a Sonnet model.
Frank 33:47
Mm-hmm.
James 33:47
said I want you to review. We have the context of the PRs and then battle each other. ah Come up with your own decision, then battle and decide together and fight for with each other.
Frank 33:53
Oh geez. Yeah.
James 33:58
like and Then give me an impact.
Frank 33:58
That's some pro level there.
James 34:01
And you don't have to go that far. If you just do the the review agent, you can say, oh, okay, yes, yes, yes, no. But what I also like to do is this. I'll do um i'll do a hybrid, which is I will do a the pull request, and then I'll do the review. So I'll do the code review. And then I won't review the code review review. I'll pull it. I'll say, Hey, listen, on this pull request, I ran the code review and this, and and it left a bunch of issues. Can you go validate all the issues and give me a report of what you think is actually important based on the code review?
Because I think about it as this.
Frank 34:32
Yeah. Yeah.
James 34:34
I don't like to think about these AIs as humans, but i like to think of them as personas, right? And a persona is like, this is a code reviewer. Frank could be reviewing my code. Heather could reviewing be reviewing my code and or the GitHub Copilot could be reviewing my code. Do I blindly trust everything that Frank tells me or Heather tells me or the code review agent tells me?
Frank 34:50
yeah
James 34:54
No, I need another set of eyes. They're either my eyes or another pair of eyes on it, right? So Frank reviews my code's least comments, I don't blindly trust them. I might have Heather reviewed the comments and my code as well.
Frank 35:05
Yeah.
James 35:06
So that is like a nice approach. And I think we forget sometimes in this world of AI of how we work in software development. You can apply those same concepts to working with AI, right? At the same time, you know, and I do review the code, but also like have another set of eyes. It's there, it's waiting for you, right? To review those, that review.
Frank 35:26
yeah
James 35:28
And I think that has been really helpful. And I've caught some really be intriguing bugs. And I would also encourage you to, to do this in your code, which is like, do a performance analysis, do a security audit, do what, like tell it to do the things that you should probably do and then have it create detailed, then do deep research on each of them.
Frank 35:42
Yeah.
James 35:48
Right. And then generate reports for it.
Frank 35:49
Yeah.
James 35:50
Right. I did this thing, which was, I use, um, um, it'd be really crazy if this actually, if it, if it happened, but we have, um, of automation for the the GitHub Copilot CLI, um the Twitter account for the changelog. So whenever it gets the changelog, it sucks up the changelog. And what it does is it does an AI generator report. So it takes all of it and it shoves it into ah Microsoft Foundry and I have it generate a really cool report with the top features and categorize them into all this stuff and add emojis where appropriate.
And one of the reviews was like, hey, you like you're not doing any like prompt injection detection. And I was like, well, I probably don't need to use it from the change log. But I was like, but I probably should? Like, I probably should, right?
Frank 36:40
I that's Twitter. Yeah, anything could happen.
James 36:43
And I was like, okay, cool. Yeah, do that, right? So that was a nice one. it was like security audit. and it's like, oh, I see that you're calling an AI service. The thing that you probably want to do are these things. So it was neat a neat way of thinking about just using these agents in different ways instead of like, go do this thing, right? Oh, think about this thing in a different way. And I've been doing this other thing, which is using these tools together, which is like do deep research in this thing and then take that output and then put it in another thing, right?
Frank 37:01
Yeah.
James 37:06
Just like how you're doing it. Like take this research, put it in an issue, then assign it to a thing, right?
Frank 37:07
yeah Yeah, I think I will give that a whirl. um Because my biggest complaint through the process, I really enjoyed my process of creating an issue and then a PR. I i feel like we got good quality code out of that usually.
James 37:24
Hmm.
Frank 37:24
um But I was still catching stuff in the PR and it was getting a little tiring hand reviewing all the code.
James 37:30
Yeah.
Frank 37:30
Oh, what a tough life I live in 2026 having to hand review all this code. um So yeah, I think I'll give that a whirl though of trying to set some agents to do the review first. I did notice in Copilot on GitHub, it does run some kind of sub-agents where it does like at least a little bit of a security check and something, but it's a bit black box. like I have no idea what's actually going on there, but if I had my own agents with my own prompt and saying something, yeah, I could see how that would be beneficial. Worth a try, worth a try. I did get a little tired of like clicking all the boxes in GitHub because like with every PR, you know, you you got to wait for it to finish, which is a random amount of time. It can be two minutes. It can be 42 minutes. You know, you never know how long it's going to take. So I wish there was a better notification system there.
And then you got to click the box that says ready for review. Then you got to go review, fill in all that stuff.
James 38:26
Yeah. Yeah.
Frank 38:28
Then you have to approve the workflows. Then if you had any changes requested, you got to repeat all those steps all over again.
James 38:30
Yeah.
Frank 38:35
um So I feel like the GitHub UI can be improved for this new brave world that we're in in 2026. And I know, I know there's there's different code editors out there and everything. But I like working through the GitHub UI because it's not on my machine. um It's in parallel. It's doing its own branches. I'm not filling my beautiful directory structure up with gross, gross work trees. So I just like the cleanliness of it all just being up in the cloud.
um But I feel like the UI could use some refinement.
James 39:10
Yeah.
Frank 39:13
I was that close to vibe coding an interface because all this stuff is obviously accessible through the GitHub APIs. So was like, very close because I was getting tired of clicking all the buttons to streamlining the experience. And I might still. We'll see. We'll see if i'll see if GitHub decides ever to update their UI ever again. i feel like it's been decades since they've updated their UI.
James 39:42
that thing's happening. think per perf I think is the first before facelift.
Frank 39:43
We'll see.
James 39:45
So you know,
Frank 39:47
It's not face. It's just, it's UX. It's how many clicks do I have to do for this thing?
James 39:51
it is it's UX. Yeah, I agree.
Frank 39:53
Yeah.
James 39:53
I do a lot of that stuff from the mobile app and the mobile app I to do the same thing I have to boom, boom, review, boom, click merge, go five too many clicks.
Frank 40:00
Yeah.
James 40:02
Yeah, so we'll see if that changes.
Frank 40:02
Yeah. yeah
James 40:04
And and that that is like an anti-agent flow, right? it's like very manual, like lots of safety checks. um So let me be introduce a few of that evolves. But it seems like you're evolving slowly but surely, Frank. I love it. It's good.
Frank 40:17
Not so slowly, man. That was high efficiency the last few days. We'll see if it's a burnout, though. We'll see if, like, can I keep this up for five working days in a week? Because feel like this high productivity is going to come at a cost. I'm curious to see how we all feel in, like, two years. We're going to be like, man, I generated so much code.
James 40:37
Well, it sounds like he works really hard for two days and he's gone vacation. Done. Shipped it.
Frank 40:42
I gotta pull a James and work from the airport. Get the app installed and just do the clickety-click in the airport.
James 40:49
like it. All right. All right. We'll go deploy your own fleet of agents on your code base. have any tips and tricks, write into the show, mergeconflict.fm or leave most likely a comment on our YouTube. Now, if you haven't caught last week's pod, I posted it late, Frank. Oops, it's on Egentic Workflow.
Frank 41:05
Oh.
James 41:06
So, you know,
Frank 41:07
I didn't notice that.
James 41:11
if you go and check it out, go check it out. And it's got really nice. It's got a really nice camera lighting.
Frank 41:25
Lighting? Camera.
James 41:26
So very nice, nice nice camera lighting.
Frank 41:26
It is. We had a red camera.
James 41:29
So give it a look.
Frank 41:29
Yeah. We should have made a movie while we were there.
James 41:30
We got a red camera. It's pretty good, it's pretty good. So give it a look um and see what you got. And
Frank 41:40
Is that better than an iPhone camera?
James 41:44
yes, I think it's tiny tiny is bed better.
Frank 41:44
Hard to say.
James 41:48
it' I mean, it is hard to say, but it's tiny is better.
Frank 41:49
Okay.
James 41:51
We're both in the studio together, so give it a look-sies if you're listening to this. Obviously, if you're listening to the last one, youtube.com forward slash at Merch Conflict That's going to it for this week's Merge Conflict, Frank. So until next time on James Montemagno.
Frank 42:06
And I'm Frank Kruger. Thanks for watching and listening.
James 42:10
Peace.