Skip to content
Episode 501

501: Autopilot, Fleets, and Parallel Agents Explained

February 9th, 2026

35 mins 57 secs

Your Hosts

About this Episode

In this episode James and Frank walk through the latest Copilot CLI power-ups—Autopilot loops, experimental Fleet/parallel agents, and Opus model/context updates—while demoing how they used plan mode to spin up a full MAUI pet‑insulin app end-to-end. Learn what Autopilot and Fleet actually do, how parallel agents orchestrate work, plus practical tips (watch your context window, use plan mode) for turning AI agents into fast prototypes.

Follow Us

⭐⭐ Review Us ⭐⭐

Machine transcription available on http://mergeconflict.fm

Support Merge Conflict

Transcript

James 00:00

Welcome back, everyone, to Merch Conflict, your weekly developer podcast, talking about all things in the world of software development, AI, LLMs, and all those goodies for developers. I'm one of your hosts, James Montemagno, and with me, the one, the only, Mr. Frank, don't know his middle name, Kruger.

Frank 00:17

After all these years, after 500 episodes, you still don't know my middle name. Hi, James. It's wonderful to see you. And I don't know your middle name either. So what what the heck am I talking about? You told it to me once, though. I just have a bad memory. I'm going to guess it's Alan.

James 00:32

No, your middle name is A.

Read Full Transcript Hide Full Transcript

Frank 00:37

Dot.

James 00:38

A dot. Because your email is like F-A-K, so it has to be Frank A.

Frank 00:39

Yes.

James 00:44

Kruger. My initial J-G-M, so wouldn't be A.

Frank 00:45

You're on to me.

James 00:49

It would be something very Italian.

Frank 00:51

I think it's just G dot or Giorgio.

James 00:54

G dot. Giorgio. No, it's Gaetano. Gaetano.

Frank 00:58

Gaiatano.

James 00:59

Gaetano.

Frank 00:59

Cool. This is what people tune in for is us learning our middle names together.

James 01:01

I'm Italian. got it.

Frank 01:04

It's good to see you, buddy.

James 01:05

Uh, no, let me see how, see how yellow I can get it.

Frank 01:08

Episode 501. We're past the hump. It's all down the hill and up the slope and over the river through the woods.

James 01:14

Bro. I could not, I could not fathom that we did an entire like live stream and podcast last, last week, randomly. People didn't know, but I was completely unprepared for everything. Um, we,

Frank 01:28

I didn't notice.

James 01:30

Uh, we just showed up. I saw that Frank had who tweeted that it was episode five hundo. And i was like, oh crap. I, I told, I told Frank, I told him, I told him, I told him when do like a live stream, do a whole thing. I set up a form.

Frank 01:42

Keep it quiet. Don't tell people. oh wait, no, you said the opposite. Tell everyone. Get everyone's feedback.

James 01:49

Tell the people, tell the the people need to know. um that was really good. Thanks for a few people that tuned in and obviously everyone that's been watching and hanging out really appreciate it Um,

Frank 01:58

Yeah.

James 01:59

Yeah, it's been it's crazy to think that you know we're halfway to 1,000 and it's only been Mm-hmm. Mm-hmm. you know ten years stuff

Frank 02:06

20 years. How does time, and my does time fly, doesn't it, James? I don't know. Like we said on that episode, we've covered so many topics that every week I just wonder what topics could possibly be left in the realm of mobile app development, AI development, and basically middle name discovery because these are the hot topics. Yeah.

James 02:29

They're hot. Well, let's talk about two things today. Let's talk about some AI models and some things I've been building. And I also want to talk about some new LLM stuff, um some neural network stuff that you're up to.

Frank 02:35

Ooh. Ooh. Oh.

James 02:41

I think those are similar. ah I want to talk about the the Opus because it came out. It's available in VS Code. There's three models. ah Opus 4.6, Opus million token, and Opus fast. um I believe Opus 4.6 and Opus 4.6 fast are available inside of VS Code and the CLI. um i don't know about the 1 million token there.

Frank 03:05

Yes.

James 03:08

i have I have models.

Frank 03:09

I have not gotten. i have not i have never heard of this thing. i did i i thought we were already getting a million tokens. That's a new thing?

James 03:18

Well, it's the context window. So they I believe that they released the 4.6 and on the same time, same time like they have a model that's like 1 million

Frank 03:21

Yeah.

James 03:28

token context window if you really need to blow up that context of window.

Frank 03:29

OK. There is. It's important, um you know, like um they put a new UI feature into VS Code. We're both VS Code shills for whatever reason. I know your reason. I don't know why I'm. But um they just did a nice update.

James 03:45

it's good.

Frank 03:45

it Yeah, that's why. um They did an update with the UI where there's a little pie chart that shows you you progressively filling up your context window.

James 03:50

Oh, yeah. Yep.

Frank 03:55

And it's a little bit stress-inducing, I have to say. i don't like to see too much pie in my pie chart, but it's there. um And then you start to discover, oh, maybe I shouldn't have a thousand different tools selected because they are taking up something like 30% of my context window.

James 04:02

Yeah.

Frank 04:11

Maybe I should back off on the 8 billion tools.

James 04:12

Yeah.

Frank 04:15

I didn't realize how like VS Code extensions can just inject tools and you have to kind of manually go turn those off. So, PSA announcement, everyone, now that they have a UI to show you how much your tools are taking up with the context window, maybe pare it down a little. you know You don't have to include every tool with every chat session and all that. But I totally get why a million would be important because I think most of the context windows I see are like 128K, which is huge, huge, but not as huge as a million.

James 04:40

Yeah. Yeah, I was doing ah a whole bunch of work and it took a while to get to that threshold, but i'm totally you're totally right right there.

Frank 04:50

Mm-hmm.

James 04:52

like Checking what you have, checking like what you're doing, XYZ, planning, doing stuff. Obviously, it auto compacts and a refresh automatically.

Frank 04:59

Yeah.

James 04:59

So even if it gets to red, it's there. I do want to point out, if you go into themes, so hit that little gear on the bottom left, Frank, if you're in VS Code.

Frank 05:07

Ooh. Mm-hmm. Mm-hmm.

James 05:08

Themes, ah color theme, I guess that is command KT. ah there's a new There's two new themes, but there's one, i believe it's at the bottom, and it is going to be called the VS Code experimental dark theme. It's a new dark theme, and there's a light theme, but it says experimental.

Frank 05:27

and Okay.

James 05:30

It should be in there.

Frank 05:30

Looking, looking. bad at reading. Oh, I see it.

James 05:32

It's at the bottom.

Frank 05:33

VS Code Light Theme Experimental. the VS Code yeah Dark Experimental.

James 05:36

Keep going. It's pretty good.

Frank 05:39

Heck yeah.

James 05:40

Would recommend. ah Give it.

Frank 05:42

Oh, more gradients.

James 05:42

Yep. Mm-hmm. Mm-hmm.

Frank 05:44

I'll say that's my initial feedback, more gradients. The green has changed. i have to decide if I like It's kind of a minty green. Let's see if I'm in a minty mood. But I see what they're going for. They're going for the cyberpunk synthwave vaporwave neon colors on a dark and misty rainy street alley.

James 06:05

I like it. You know, any visual or refresh that's in the box, you know, is really nice to have. I'm usually just a default. Whatever's there is there. So this is, I believe will probably become the new default at some point, but looking for feedback.

Frank 06:12

Mm-hmm. OK.

James 06:15

So dog food it, give it a whirl and let them know what you think. But I think that a nice little hidden feature in this release. Tons of little hidden features. In fact, you know, if you go into, going to give you a hidden feature, or go into settings and then, oh, well, my settings are all fancy pop-up because i'm on insiders. um Type in work

Frank 06:33

ha

James 06:35

bench I believe is what we want.

Frank 06:38

Ah, uh-huh, uh-huh, workbench, okay, yep only 159 settings found.

James 06:42

ah Maybe it's on welcome. type Type in welcome. Yeah, there you go.

Frank 06:51

yeah.

James 06:51

And you want workbench startup editor.

Frank 06:55

Okay, yep, welcome page, sure, yup

James 06:57

Yep. Yep. Do agent sessions welcome page. Okay.

Frank 07:01

Oh, boy. Are the agents going be talking to me as I wake up in the morning? We shall see. We shall see.

James 07:09

and then I don't know how to bring it back.

Frank 07:12

Shift command end.

James 07:12

i don't know What is it?

Frank 07:14

Oh, yeah. ah Just to get a new window. Yeah.

James 07:18

I think you just say welcome, right?

Frank 07:21

It's cute. It's here. it um So the new welcome page in VS Code ah puts up a classic little chat window.

James 07:24

No.

Frank 07:29

And I'm wondering where this chat history is from. Is this the co-pilot chat history? Hard to say.

James 07:36

It's a bunch of my different repos that are in here. So I believe it is a bunch of different repos between like your machine, other machines, things like that, that are coming up. Yeah.

Frank 07:47

I have mixed feelings because I like this. I want to be clear. I like this. But I'm missing my recent file list. Why can't we have both?

James 07:53

Hmm.

Frank 07:55

Can't I have my cake and eat it too? Because I'm a little... One click is always better than two clicks. Just saying if you're ever debating a UI thing, one click better than two.

James 08:03

Hmm.

Frank 08:07

And it's better than three. Far better than four.

James 08:11

Yeah, i think there's I've talked to the team a lot. I think there's gonna be a lot of things happening here. i'm pretty excited about it in general. you're see some really interesting things. I do agree with that. That's a good feedback here. There should be the recent folders, et cetera, et cetera in here. That is for sure.

Frank 08:25

Yeah. Ooh.

James 08:27

um But I do like with some sessions. Are you opening up fresh? Good to go. um Could be neat. Could be neat. Okay, um okay so I built some i built some stuff. ah Frank four, six came out and I said, why not give it a whirl? I didn't do my traditional do a thing, but I wanted to um experiment with the CLI a little bit.

Frank 08:41

that it

James 08:46

I want to point about the copilot CLI. There's a few new features that I want to point out here. There is a new feature um called autopilot.

Frank 08:58

Okay.

James 08:58

And this is a Ralph loop built in to the CLI.

Frank 09:02

Nice. You know I'm a fan.

James 09:03

Okay. and fan

Frank 09:04

Mm-hmm.

James 09:05

And then additionally,

Frank 09:06

I got to be honest. I haven't actually used a Ralph Fluid book. I'm talking about it from a theoretical standpoint. I just like it because it's an AI just kind of running on its own. That's the part that I like the most about the Ralph Fluid.

James 09:15

Yeah. Yeah, this one for all intents and purposes seems like you give it ah a plan or a task, give it the loop and then it just kind of goes to town. ah

Frank 09:27

Mm-hmm.

James 09:29

And then there's another new feature that they just added called fleet. ah So slash fleet is experimental. And, but you can just like turn on experimental. So what fleet does is it will take ah whatever you give it, break it into tasks, store it into like a SQLite database, and it will deploy a sub-agent fleet to work on those tasks asynchronously

Frank 09:52

Okay.

James 09:55

And then, of course, this is all working in autopilot, which the best way to work in is a YOLO mode, which just auto-approves everything.

Frank 10:01

Oh, gosh, yeah.

James 10:02

And hopefully you're in a container.

Frank 10:02

Whew.

James 10:03

I'm not. But just go to town. um So you then have this rough loop with sub-agents that are working all asynchronously on tasks, and they can check in on each other to see if there's going to be conflicting paradigms that are going on. So what I did is...

Frank 10:19

Fascinating.

James 10:21

If you remember my pet insulin app that I built was GitHub Spark a long time ago.

Frank 10:24

isn

James 10:26

I wanted to make it reality into a real mobile app. So I booted up a blank Maui project. I said, Maui, let's go, baby, with four six. You got the knowledge. It was go time. And I wrote up a big um ah big like paragraph of what I want the app to be. I put it into plan mode in the CLI. I said, let's go, baby. So a big old plan mode. Go to town. I said, let's plan it out. answered all the questions, everything it wanted. I said, I wanted a front end. I wanted a backend. Um, I wanted a promotional website. I want these fancy graphics.

want all this stuff. Bump, bup, bup, bup, bup.

Frank 11:00

Big plan.

James 11:00

Ask me bunch of questions.

Frank 11:01

OK.

James 11:02

Yeah. And I said, I said, also do tons of research on like pet insulin, right? Difference between cats and dogs, all the different types of insulin you could get all this, all this, you know, give me all the deep details.

Frank 11:09

Hmm.

James 11:14

And I sent a big plan that came up to it. And I said, autopilot fleet, let's go. Just let it rip. And I said, let's, let's four, six it. And, uh, it pretty much, ah here's the interesting part is it's not even one shotting anymore. Cause a Ralph loop inherently isn't even one shotting it, right? It is, but it's doing a whole shebang. The crazy, the crazy thing about this is as I watched it go and checked in on itself and compiled and looked it,

Frank 11:36

Yeah. Mm-hmm.

James 11:47

wanted super clean code. And when it was done, it took like 30 minutes or so to do this crazy thing. I'm talking Azure Functions backends. I'm talking like the whole front end, promotional marketing website, like the whole shebang. um And I was just watching TV while i just did. I wasn't paying really that much attention. I was like let's go, baby.

Frank 12:07

Oh

James 12:07

And it was just going to town. But as I was watching it, it was not happy with any warnings in the code. It's like, I refuse to have warnings.

Frank 12:16

gosh. Yeah.

James 12:18

So it like, by the end of it, it gave me flawless, zero warning, obviously zero error, anything like that. It pulled in SQLite-net, it pulled in all this stuff, it did everything.

Frank 12:27

Nice.

James 12:28

And it was great. And and it and it just used out of the box control, so nothing fancy. And it was a whole onboarding flow. And I said, this looks great, but let's go super fancy. I want themes. Let's use some Skia Sharp. Let's do some custom gradients. let'ss like I want it whole shebang, the whole thing. And I want the the website to do all this stuff. gave it an another loop, put into plan mode, 10 minutes later, boom, this beautiful app. It has like a bunch of themes. You slide there, it has this beautiful like welcome thing like that, like walk-ums, adds a pet, this whole thing.

It has share codes, so no no authentication required. Everything's a share code because you know me, share codes only.

Frank 13:02

Yeah.

James 13:03

um You can have a guest that has unlimited access. You can have a guest that has like, um you know, read only acts and not read only, but only like a write only access so they can do logging. So if I'm to a pet sitter, they can't see all the logs, but they can post logs.

Frank 13:15

Right, right. Got it.

James 13:19

Um, haven't figured out like there's some authentication stuff that I probably want to do, like revoking some things.

Frank 13:24

Yeah. Yeah.

James 13:24

Um, but, I was really, really impressed. Like the whole website was done, everything. It's on my GitHub Pet Insulin Tracker or whatever. um even came up with some names. It did a whole thing and it was kind of crazy. So ah I was really impressed by it. Now, I don't necessarily know if that was 4.6, right? If 4.5 would have done the same thing or not, right? It's really hard for me to know at this point on these models, but at least whatever that loop was and whatever it did, i didn't have to worry about context.

didn't do anything. And the other cool part is like these sub-agents, and have a whole podcast that's come out the same day this podcast comes out for the VS Code podcast with Harold, talking about sub-agents that are all like in their own little tiny context window doing just specific tasks. So that doesn't bloat your main context or just like little tiny ones. But I was super impressed.

Frank 14:13

Yeah.

James 14:15

I haven't used the 1 million. I haven't used the fast model at all yet. Not that sweet, sweet 9X multiplier. But um I thought it was pretty good. And and I will say like, I'm a big fan of using different models for different things. You know what I mean? Like I'm not a big fan of just sticking to one model because inherently Opus even sonnet, you know, they could be a little slow, like a GPT five, two, and the codex models just came out with an inherent fast mode that just everybody gets him.

Bro, like, dude, it's fast.

Frank 14:46

Nice. Yeah.

James 14:49

It's like rips. um So I do a lot of planning and with those models specifically and get the plan really, really good. And then I still kind of like, so I just, I kind of gel with Sonnet really well, but I ah gel really good with like 5.2 or 5.2 codecs, especially if I know my code base really well. I think that's been really good. But I've just been trying it out, seeing what it looks like. And I was really impressed by what it has. Is it perfect? No, just like any app that we know, if I want to actually production it, I'm going to spend a bunch of hours really massaging it, testing it, doing all this stuff.

Frank 15:20

Okay.

James 15:20

But for a prototype for a Maui app with the back end, with the front end, deployment steps, CI, CD, I was like, all right, I see you, bro.

Frank 15:21

Yeah. Yeah.

James 15:30

Like pretty good.

Frank 15:31

Okay.

James 15:31

Pretty good.

Frank 15:32

Dang. Okay. let's Let's break down a few things because i just i'm I'm a total noob. I haven't seen your other podcast, so I can't believe you're cheating on me with another podcast.

James 15:42

Sorry.

Frank 15:42

But um I am curious. The big problem with doing the parallel sub-agent thing is keeping them from interfering with each other, having them work on the same task and all that kind of stuff. So are we talking work trees, as we discussed? You said you're not containerizing, so that's not keeping them separate. Are they really just bashing away on the same code and you're just hoping for the rough loop to fix it all?

what's What's the deal with, or or are they not actually parallel?

James 16:12

The fleet is parallel. There's multiple running at the same time.

Frank 16:13

Yeah. Yeah.

James 16:15

I do not know if they're working in their own work trees.

Frank 16:17

yeah Good answer.

James 16:20

They may be working in their own work trees and merging things together. But I do believe that there's sort of like an orchestrator that is figuring out, hey, what's the potential of this? So for example, if if um if it's like what i what I noticed was that a bunch of them spun up when I was like, hey,

Frank 16:31

Yeah.

James 16:40

I want you to do this on the front end, do this on the back end. And then I gave it a big task list of things to do. and it's like, cool, I can literally do a sub-agent that's like only working on the website because who cares, right? The website has no dependencies.

Frank 16:54

okay

James 16:55

So I think the orchestrator is analyzing the task list and saying, I think it's probably being pretty careful. like What do we think there will be content you know things there back and forth? But it might be working in work trees inherently. I'm not positive on it. I'd have to check with the team. But that'd be pretty cool. I did notice that the new work tree support in the latest VS Code is like crazy good.

Like when you spin off a background agent, you can just hit apply and it'll apply it right away. like you don't to worry about merging and things like that. just like bring things right into your main line. Pretty cool. But I'm not sure if that's how this is working. i have to like look at it. i just have to watch the, you know, is there a work tree directory, right?

Frank 17:31

Yeah. Yeah.

James 17:33

But I know that there is like a tracking mechanism and what it could be doing and what it should be doing probably I'd have to ask the team is going off, analyzing each task first and saying, okay, what files is it probably going to like almost a mini plan for each task and then doing stuff and then deploying.

Frank 17:50

yeah

James 17:52

Cause what it has to do is it not only has to look at all the tasks, but as to order them, right? Like you can't update the backend until the DTOs are updated and like, you know, things like that. But I did see them light up.

Frank 18:03

Yeah.

James 18:05

Like I wasn't getting like 20 parallel agents. I was getting like two or three, you know what I mean? Like often two.

Frank 18:10

Yeah, so there' okay so there's something in the background ostensibly trying to create some kind of ordering to all this kind of stuff. But it it is fundamentally a chicken and egg problem. like you You don't know what order or what files it's going to touch until it starts working. That's just how these networks work. you know like if If you could predict the end result, then you could plan perfectly, but you can't predict the end result.

James 18:29

yeah

Frank 18:34

You have to execute it to do it. So anyway. That's just me kind of thinking out loud. So I definitely want to play with this because I want to go back to the plan thing. So I'm curious. I've been actually doing plan mode, James, lately.

James 18:48

Good.

Frank 18:49

I'm like, you know, there's this whole other mode.

James 18:49

Every day.

Frank 18:51

Maybe I should try it sometime. And I've really gotten the religion of starting in plan mode. And I really wasn't doing that before this week, to be thoroughly honest. I'm just like, YOLO, let's just write some code. Why... Why would I plan? That sounds absurd. um But I've actually been enjoying the plan mode mostly because it asks me questions. And they're usually questions that I didn't really think of ahead of time. And that's kind of the best. I like it when it's prompting. me, the human, the AI is prompting me. And I respond in my own chat window. So i'm curious, um inevitably, in the plans, and I'm surprised you did such a big plan for like a total app with all these components and everything. Did it break it down into like a list of 12 things there, and then when you did your fleet, you ended up with 12 sub-agents?

Or was the plan... you know I'm just making up numbers here. Or was the plan like five items and it created 15 sub-agents? Or was it pretty strict to sticking with your plan? I guess that's just the general feel I'm asking for. Mm-hmm.

James 19:55

is pretty strict to following with it in general. I'd have to kind of look, that I have the whole um gist.

Frank 19:58

Yeah.

James 20:03

So you can do slash share and it's kind of hard to parse like real time, but you can see like ask user, create, and then I think you you'd be looking for is look for a plan summary.

Frank 20:03

Yeah. Mm-hmm.

James 20:16

Now, I also will say, I don't know if I use the fleet right away. I might have just gone into Autopilot and Autopilot will also do the things as well. But you can look at the plan. It's 11 phase, 20 work items covering the full app. um I don't where the 20 work items are, something in there. But I think it doesn't give you every single bit bit of detail. But inside, then you can see all the asks. You can see what it asked me, the arguments, XYZ, all these things. And it kind of went to town. So it's creating like a SQL database, SQLite database behind the scenes.

And then just starts to get to work on it. And it's hard, you won't be able to see it in this view, like what it was doing. Like if it was a sub agent, if it was a fleet, but that doesn't show up in here.

Frank 20:51

Yeah. Yeah. Mm-hmm.

James 20:54

That would be kind of cool. Maybe the team will add that. i don't know if it's in the logs. It just kind of happens. But what I saw basically was in the UI, you will see the main agent running and then you'll see another, a little little counter that says background agents working. And that background agent working ah was bopping between like one and two and zero. Like not a lot, but a little bit where it was like, oh, I can go delegate this work off over here.

And, you know, I think where this would be really important is like, hey, okay, I want to create like the the plan. I want to come up with an implementation. okay Then I want to come up with the tests and, like imagine you're doing bunch, like you could have different sub agents that are, you know, doing each of those little bits and pieces of it. That's what I imagining it doing. But um it seems pretty cool. And they're, they're, they're not even sub agents, they're parallel agents. So I'm probably saying that correct and incorrect.

Frank 21:49

Right. That's the important thing. Yeah.

James 21:51

So sub agents are really like read only parallel agents are multiple agents running at the same time. So let me give your main agent that orchestrates multiple sub-agents, which are just gathering context, returning it back. But in this case, the main agent is orchestrating multiple agents themselves that are going off and doing things, which I think is very fascinating to go off and do.

Frank 22:10

Yeah.

James 22:14

Yeah.

Frank 22:16

So it's a super agent. Agents, agents, agents. went Back to me complaining about using that word for everything.

James 22:19

a Agents all the way down. So, yeah.

Frank 22:23

<unk>s In this case, it's true.

James 22:25

Yeah.

Frank 22:25

um Yeah. Okay. this This sounds like I'm going to burn through all my credits in the month very quickly, um but i'm like I'm a little bit excited.

James 22:31

Good. It's a short month.

Frank 22:35

It's a short month. I like that. Thank you. Thank you for the optimism. Got to burn through those credits somehow. I feel like I will give this one a try. um The parallel, I'm not 100% sold on. I want to see that work, but I am very curious about the autopilot. Yeah. you know the The biggest thing I've seen with the Ralph loop is you do the autopilot, especially in the YOLO mode where you don't have to give permissions all the time, so that it like even if it fails to solve the problem, it'll kind of pick it up on the next loop around.

I'm curious to see how that works. so I hope some of this stuff is documented. I know they don't document everything in detail, but I hope they give some rough idea of how the autopilot was it. algorithm works, copilot, autopilot.

James 23:21

Autopilot.

Frank 23:22

Makes sense, yeah. And then the fleet. Yeah, so I'm looking forward to reading the docs on it all to understand it better. You got an app out of it, though. How hard could it be?

James 23:33

Yeah, if I can do it, anyone can do it. No, I think, and I'm i'm still torn, right, because I'm inherently not a CLI person, but I've been like testing a lot of things out recently just to see how they work.

Frank 23:42

Right.

James 23:43

And I still find myself very much back and VS Code because I really like viewing files and viewing changes and seeing things happen in real time.

Frank 23:52

Yeah.

James 23:54

Not everyone's that way, but I i like to, like, i even in the CLI, I'm watching what it's doing.

Frank 23:55

Hey.

James 24:02

Like, I'm, like, looking at it, but you can't

Frank 24:03

Mm-hmm.

James 24:04

like see the code change. and I will say, I think there's a VS Code extension that Damien wrote from the Aspire team that will like automatically like open files and show the diffs in VS Code if you run it from inside or whatever. It's like autofollow or whatever.

Frank 24:19

Mm-hmm.

James 24:20

So to auto-follow what the CLI is doing, which is cool.

Frank 24:21

Yeah.

James 24:24

But I know that the team is like working on some integrations there, I think.

Frank 24:25

OK.

James 24:27

So I think that would be better. But then I'm in VS Code. So you know might well just use VS Code.

Frank 24:33

Yeah.

James 24:35

But I really like seeing files. i like seeing the code. you know It's important to me.

Frank 24:38

Yeah.

James 24:38

Especially in more mature projects, like maybe in this first one, I'm just like, let's see what happens. YOLO, right? But like when I'm pretty serious about the code, like I'm just deploying for the Switch website, like a full refactor where I'm using a SQLite database and I'm like doing all this like migration and syncing and all these other things.

Frank 24:44

Mm-hmm.

James 24:55

I'm like, I really want to look at the code when I'm asking you to do a thing. Because when I see your output, I'm like, nope, that's not actually what I wanted. What I wanted was this. And it's like, oh, you're absolutely right. that We should have gone this route, right? and i'm like because i And I know I could see that, but I feel like with the CLI, it's inherently like not fire and forget, but like fire and like come back later and see what happened, right?

And and and then review it.

Frank 25:20

Yeah.

James 25:21

And I'm just going to open up VS Code to do that. So I'm a little torn, but...

Frank 25:25

Mm-hmm.

James 25:26

the CLI does update like five times a day. So it's like fun to see the team crank away. i mean, granted like VS Code updates every day, but I'm like, I got four updates every day. So that's neat. So yeah.

Frank 25:38

OK. Yeah, and I'm glad you clarified that. These are all features of the CLI and not necessarily in VS Code yet.

James 25:43

Yeah.

Frank 25:45

yeah I hate that we're forked a little bit.

James 25:47

Yeah. Oh, talking of the CLI. um Mr. Steve Sanderson himself, they just released an update to the GitHub Copilot SDK, which we talked about previously on the pod.

Frank 25:56

Okay. h

James 25:58

ah There's tons of updates, but the biggest update is it now, i don't know if it's optional or if it just automatically does it, but it will bundle the CLI into your app automatically for you and then run it and deploy it.

Frank 26:07

All right.

James 26:12

So it doesn't even have to be installed on the machine. It'll just work, I guess. I don't know how that works, but like awesome.

Frank 26:16

this This blew my mind a little bit um because the CLI, at least for me, like I'm on a Mac, so it's like Brew install the CLI and it's it's not very big and and it just kind of goes somewhere wherever the heck Brew puts all its junk and and and then it's there.

James 26:31

Yeah.

Frank 26:32

And now they're just like, oh, F it. We're just going to throw the whole thing into the package.

James 26:36

ah Let's go.

Frank 26:40

YOLO. They're definitely living that life.

James 26:41

Yolo.

Frank 26:42

um Yeah. Yeah.

James 26:44

Because that that way you're versions of the CLI always match, which is your problem early on.

Frank 26:44

just Yeah. It makes sense. It makes sense. But then like I'm like, is it a CLI anymore?

James 26:52

Yeah.

Frank 26:54

is it is it isn it Isn't it just a library?

James 26:56

it's it's a It's a runtime, it's like a runtime, right?

Frank 26:59

Yeah.

James 26:59

That's inside of it. I will say, that's true.

Frank 27:01

It's an agent that we call a CLI for whatever reason.

James 27:04

It's powered by the CLI, which is now just an agent. Yeah, and maybe they'll change it, who knows.

Frank 27:08

Yeah.

James 27:09

I will say I've been using my, know they have NPM and you have like pete whatever the Python one is where you can just like easily pull down a package.

Frank 27:15

Yeah. Yeah.

James 27:17

I've been using this DNX thing all the time, the Nougat version, the.NET 10 version.

Frank 27:20

Nice. Yeah.

James 27:22

And boy, it's so nice because it'll tell you if there's an update, they'll tell you like this when you run it. I do, I do DNX podcast, metadata generator, boom. It's like right there, sucks it down any machine.

Frank 27:33

Yeah.

James 27:35

i just got to remember the name of my packages and my apps.

Frank 27:38

Yeah.

James 27:40

But this, you know, here's what's neat, dude. Like, um, You know when you got to release an app like onto machines? Like, I mean, granted, there's a GUI, but like it's really complicated.

Frank 27:48

yeah

James 27:52

Like, you know, DNX it. Do just d insult the Nougat.

Frank 27:56

It's DNI.

James 27:56

Just run it. it You new need to sign it because it's already.NET's there. Genius. So good.

Frank 28:03

WinForms. So are we going to do our WinForms apps with DNX? can i Can I ship a Cocoa app with DNX? I think we need to test all these GUI things. Because you know we you were talking about the CLI, and you're like, I kind of prefer VS Code. i' like, yes, James, windowing environments are good. There's a whole reason we moved away from stupid term um environments. So but you know whatever. They've got some cache these days. They're in vogue.

James 28:26

I mean, I would say like if you can NPX something, just like a node app, you should be able to NPX like a Blazor server app that's just like a package.

Frank 28:31

Yeah.

James 28:33

Just like run run my thing. Just do it.

Frank 28:35

Yeah. Yeah.

James 28:37

know, just run it.

Frank 28:37

don't Don't make me think.

James 28:38

There's,

Frank 28:39

we're We're living in the AI age.

James 28:39

yeah Yeah, just.

Frank 28:41

I'm not thinking here. Just run it. Don't put it in a container. No, no. Give it full access to my hard drive and let it burn.

James 28:48

Let her run. Let's go. wow

Frank 28:51

How long until there's a virus?

James 28:52

ah

Frank 28:52

and Never mind. I didn't say that.

James 28:54

Well, I'm only installing my own packages or your packages, so that's it. So, ah yeah.

Frank 28:59

It's true. We are kind of myopic in that way.

James 29:02

Yeah.

Frank 29:03

These are great updates.

James 29:04

um

Frank 29:05

um Yeah, I'm still worried about the, I love the RELF loop, but I'm still a little bit nervous about it. I'm nervous about burning through tokens and I'm nervous about it ever terminating. So I'm actually happy to hear that yours terminated in 30 minutes. Because I thought you were say like three days later it was still spinning something.

James 29:19

Yeah.

Frank 29:24

Because these are local.

James 29:24

Now.

Frank 29:25

Like you have to keep, in your case, like the CLI up and running and live. You can't just unplug the computer.

James 29:29

Yeah.

Frank 29:31

you'll i mean, Ralph can't beat you unplugging your computer. Because as far as I know, you can't do that trick where you kind of upload it. You dispatch it to the web and then it'll take over. As far as I know, that doesn't work.

James 29:43

If you just boot up a virtual machine in the cloud, then boom, that's the ultimate sandbox, baby.

Frank 29:47

Ooh, wow.

James 29:49

Yeah.

Frank 29:49

Mind it alone.

James 29:50

Boom. little dev box. I mean, that would be pretty, you know, clutch. You spin up a little like Linux VM in the cloud. Just like, I mean, I'm just saying you go access it from your phone. I'm just saying like, it's crazy.

Frank 30:03

Yeah.

James 30:05

Like the whole, I'm exhausted and excited and confused all at the same time.

Frank 30:10

Yeah.

James 30:14

Like are there's so many emotions happening, frankly.

Frank 30:18

But I like you being the beta tester for me because like i'm always I'm a few steps behind. you know I'm still enjoying the PR world of like assigning things to co-pilot.

James 30:23

Mm. Uh-huh.

Frank 30:27

Welcome to 2025, Frank. you know wait keep Keep up with all this stuff. But like I was complaining to you before the pod that I had kind of a low productivity week. i was distracted with a bunch of things. But you know what I was still able to do? Go to a website, file some issues, and assign them to co-pilot. It's embarrassing, but like I personally was not very productive this week, but I was burning through tokens and ai agents were being productive on my behalf. And I'm not...

James 30:55

they're They're just waiting for you.

Frank 30:56

i'm not

James 30:58

They're waiting for you, Frank.

Frank 30:58

there

James 30:59

They're waiting for you.

Frank 31:00

I don't think productivity is like the most important thing in the world. i think creativity and a lot of other things are more important. But at the same time, when I'm having a low productivity week, it's nice to know the robot can pick up some of the slack from me.

James 31:17

Yeah, it's just waiting for work to do. Yeah, I think that it's been a really fun time. Glad that you're still burning through some tokens, doing some stuff. um Yeah, I think also to your point earlier about the plan mode stuff, I think everyone's also in their different phase of adoption, which we've talked about in this podcast. if all these things that we've just talked about, like confuse you, like I've never heard of that, like that's okay. Like, you know, just going in, picking a model, asking it questions, doing a thing, like That's so awesome.

Frank 31:38

Yeah.

James 31:44

um you know We call it rubber ducking with it, just asking questions, pinging a thing. like It means you can go into any code base, ask it questions about it. It'll like understand the code base, do the thing. like you could you could and You could go check it out, but give you deep, deep research. i mean A great example is, Frank, we have this like shower pan that's being installed. look down into the drain I see water.

Frank 32:06

Not where I thought the conversation was going. Cool.

James 32:10

i look down into the drain and i see water And i was like, huh, so I pour some water in. like It's not rising, but i was like, why is there water there? like you know The plumbers came out, they did the thing. So I go to to Microsoft Copilot, you know like Chagy APT.

Frank 32:23

Yeah.

James 32:24

And I was like, I see this thing and I'm looking down at it. I'm give me visual indicators. I want to know if this is right. What's wrong? What should be the expected thing? And here's the whole thing. And it gave me like this beautiful, like multi image breakdown, like images from like websites, all this, pull all these information sources together. And like, could I have gone and Googled with being like all these things?

Frank 32:43

Mm-hmm.

James 32:46

Like, yeah, Totally. But it just condensed it into a nice dense thing. That's like, yeah, that's called a P-trap, you idiot. And I was like, oh yeah, you're right. Like anywhere else it's a P-trap. But I didn't realize, oh yeah, there's a P-trap underneath that. That makes sense. No gas fumes coming back. But to me, I'm just like looking down like, oh, that seems weird. There's water there. So, but then it also gave me like, here's where it may be going wrong. So it was like really cool But you can just ask it.

You can just, you can just ask it things. You can just do things. You can just build things and you can just put, like, I think that's the crazy part is like,

Frank 33:13

Mm-hmm. Mm-hmm.

James 33:18

This is why I'm so excited about it. But also like when you say, oh, I'm now using Planmo, like that makes me excited because i when I stumbled upon that moment that you had, it was like an unlock for me. Because all of a sudden, it was asking me questions that i didn't think about. It was refining things that I did think about. And the results were better with minimal time. And in the latest VS Code version,

Like the new UI, UI, baby. I love a good GUI. Give me GUI. It's visual questions with even like multi, you know, selection checks and all that stuff. It's like so good.

Frank 33:52

OK, yeah. Yeah.

James 33:53

I'm like, ah, that's, and also what I'm noticing, I haven't, I've seen it on occasion is even if you do stuff, let's say you give it a generic thing, just in normal agent mode, it will, if it runs into something, it's like, I don't know what to do. It'll ask you a question.

Frank 34:06

yeah

James 34:06

Yeah, I think it does it not that much, but like on a where it needs clarification. So I'm glad you're summing upon, if you're just summing upon these features for the first time, awesome. And if you're like, I'm not there yet, bookmark this episode 501 or whatever, and then come back to it.

Frank 34:19

yeah ah That'll be me in like two months. Two months. I'll be like, James, have you heard of this thing called Autopilot and Fleet Mode?

James 34:26

Have you heard the news?

Frank 34:29

Turns out they told me to do my job now.

James 34:32

Yeah.

Frank 34:32

Not yet. Excellent. Excellent, sir. So that's another app you've written that you're... Well, this is a rewrite. The app was written and you just decided you want a Maui one. So the other one was Spark, right? You did it in Spark?

James 34:47

Yeah.

Frank 34:48

Yeah, so...

James 34:48

Well prototype. Yeah. And that was cool. It was a little prototype to say what I wanted. I learned a lot from that too. Like what I wanted, what I didn't want it, what's helpful, what's not helpful inside of it.

Frank 34:53

Yeah. Yeah.

James 34:59

And yeah, that was good. So I'm pretty close. really want to put this on test flight so Heather and I can test it and like do it with like a dog sitter that we have. I think that would be neat. Just an app for myself really.

Frank 35:08

yeah

James 35:09

So yeah. um But yeah. All right. Well, that's going to do for this episode. Next week, I said this is going a two-parter because I said we're going talk about neural networks and then we didn't because it's 35 minutes in. So next week, Frank is going to talk about how how image generation and neural networks are creating nuclear diffusions that will never be the same again.

Frank 35:31

Oh. Oh.

James 35:34

And that we regret all of our decisions for every neural network that we've ever made in our lives. Is that correct? Did I get that right?

Frank 35:43

That's 10% true.

James 35:44

and Okay.

Frank 35:44

We'll leave it there.

James 35:47

All right. Well, that's going to do it for this week's Merge Conflict. So until next week, I'm Jason Watson-Magno.

Frank 35:52

And I'm Frank Kruger. Thanks for watching and listening.

James 35:56

Peace.

Download Transcript

Episode Comments