Opus 4.5 Changed Everything  🔗 **Source:** [Opus 4.5 Changed Everything](https://share.snipd.com/episode/b8761c64-848c-4340-ac95-a9a26c318ca7) ## Highlights - **Holiday Model Gains Plus Opus 4.5 Sparked A Surge Of Experimentation** - Adam and Burke recount how Opus 4.5 made them suddenly able to finish projects that earlier models left half-baked. - The podcast notes a holiday surge in usage (Claude credits) and a model step-change that motivated widespread experimentation. Transcript: Adam Stacoviak Well, friends, we're here with Burke Holland, a dear friend, finally on the pod. We can't tell if you've been on the podcast before, Burke, but you know what? I'm glad you're here. You work for Microsoft, working on the GitHub team, running Copilot. And, you know, I just don't know. This is a crazy world we're in. You know, what sparked this idea was actually Jared, who's since retired from the podcast. But you shared a post, one of the most recent posts, but you're not an avid blogger. Other than that, January 5th, Opus 4.5 is going to change everything. And that was January 5th. And obviously, you know, today is February 25th. A lot's changed in those two months. What do you think? Burke Holland Yeah, so just a quick rejoinder on the intro there. I work on GitHub Copilot, but that's a massive product surface area. There's the cloud coding agent, there's the CLI, and there's Visual Studio Code, and there's CodeReview, and there's the SDK. There's so many different teams working on it. I work developer relations and I work across those different groups. But so yeah, I'm not an avid blogger, am I? I think I do more video now. I do. Adam Stacoviak I do more video now. You're like, I haven't posted in a while and here comes the major. That's how I wanted to frame. You hadn't posted in about a year uh you do post quite frequently other than that but it was like a year gap and then you're like listen i hope it's four five is changing everything well Burke Holland Bam yeah i saw that and then it uh well i mean yeah that's what happened i saw that it went to uh it was on the front of of hacker news i was like that's whenever you blog this happened to me maybe Two or three times in my career and something like that happens it not the post you think, right? You just sort of throw something up. Would you have written it differently given the attention? Adam Stacoviak Of course. Burke Holland Yeah, of course. Like if you go in, there's this big like disclaimer at the top that was added after the fact. It just like is essentially everything I wish I had said coming out of the gate. But yeah, let's talk about that. I mean, I, so there's this inflection point around, right around December and everybody knows that this is true. Like I, I mean, intrinsically people can just, they just, I'm not sure everybody knows Burke. Adam Stacoviak Honestly, I think a lot of people who are the inner circles know this, but I'm having conversations. They're like, you used AI to do that. Why no one wants to use ai ai is you know this or that and you and i and a lot of people in this audience are in the know and some of them are actually like i don't know you know so maybe not everybody Burke Holland Knows not everybody knows i don't think yeah so let's let's add some context there so i've been working on github copilot for year for literally for years so i've been with it since gT-4 Which is kind of like the first model we got. And I actually created a custom chat mode called Beast Mode, which would just get 4-1 to like do things, like just to get the model to do something instead of telling you it was going to Do something and then not doing anything at all. And then the model started to get better. And then we got Sonnet 3-5 and Sonnet 3-5 was like, I mean, it was crazy, right? People were building all kinds of stuff with it. Yeah. It would actually do the thing. And it would build you a nice purple gradient website every single time. But the point is that it could do a lot. But it was a really sloppy model. It just would generate a lot of spaghetti code. Sonnet models are super eager. So they want to please you. And so they end up just doing, just running kind of willy nilly through your code base, but they will build things agentically. When we say agentically, we mean like the agent is like calling tools, it's running terminal commands, it's looking stuff up on the internet. It's acting like a developer. But the problem was, you know, even for us from where I was sitting, I could see that like it would get you to like a demo or a prototype. But then it's like you couldn't then it would get stuck on an error. And then you had a problem because it doesn't understand the error. And neither do you because you didn't write that code. And that's a that's that's a problem. You can't really sit in both spaces. You can't like have an agent that is half right half the time. You need it to be like all right all the time, ideally. And so in December, when it was November, December, I forget. Um, releases this model called Opus, which they've had Opus for a while, but they release Opus 4.5. And it's because my job is to test these things out. What does this tool do? How does this model work? I'm testing out Opus 4.5 over the Christmas break. ([Time 0:05:37](https://share.snipd.com/snip/e09baa5a-2633-4031-8cf9-40c21cd7711a)) - **Opus 4.5 Delivered A Step Function In Developer Productivity** - Opus 4.5 produced a genuine step change in capability, enabling one-shot, well-structured native app code that previously required lots of manual iteration. - Burke built multiple working native tools and an iOS app for his wife's small business in an afternoon using Opus 4.5 instead of weeks of nightly effort. Transcript: Burke Holland How does this model work? I'm testing out Opus 4.5 over the Christmas break. And the first thing that I did was I tried to, I was building like native windows tools. I tried to build a tool that would allow me to resize windows on windows to a specific width and height. Like I'm just testing models inside of visual studio code with copilot and it just one shots it just one shots it. And I was like, what the heck? And not only did it one shot it, but it used native when UI libraries and it just did it right. Right. And then I'm going through and I'm looking at the code and I'm like, this is really well structured. Right. It's not, it's not perfect, but it looks really good. And I, and the model's thinking is clear. And, and I was like, that's, that's nuts. Uh, and so then I built another tool that it was like, let's capture a part of the screen and then turn it into a gift. So like, you know, for sharing on Twitter and things like that. So I did that and it one-shotted that. And I was like, well, shoot, I wonder if I could just like turn this into a complete screen video editing tool like snag it or something like that and i did in the course of like i don't know A couple hours right i was able to build something that functioned approximately approximately not product level but it worked it worked exactly and it's like that's just, that's Just nuts. Like when you think about, and all I'm doing is prompting the model. I don't, there's no special workflows happening here. It's just me telling the model what I want, brainstorming with it a little bit, and then he goes off and does the thing. So when we went into the Christmas break, I was like, I was pretty confident that I could actually build useful things. And my wife owns a small business. Um, people who are listening to this pot to the pod may have seen this before, but like, if you're in the U S you may have driven by houses and there's like a big happy birthday sign in the Front yard. It's like different letters. Have you seen this? You may have seen this. Yeah. Yeah. Adam Stacoviak I'm in Texas. So we do things like that in Texas. We have a big front. Yeah. Burke Holland So it's a franchise, and she owns that franchise for our city in Tennessee. And so she has a lot of – What's it called? It's called Card My Yard is the name of the company. Oh, that's right, yeah. Yeah. Yeah, cardmyyard.com. Go order a sign for your – There you go. Do it up. For your kid's birthday or your dog's birthday. It doesn't matter. Okay. So she has, I build all these little apps for her. And these little like, her entire business runs on a Raspberry Pi. No way. Yeah. And it does things like, she gets orders in her email and it prints them out for her and stuff like that. But one of the things that they wanted people to do that the franchise wants people to do is to like get on Facebook and post pictures of the signs that people have put out in yards along With a caption. And like running a page for a small business is inherently marketing slop as it were, but it's like, it's a pain. People don't want to do it and she doesn't want to do it. Right. So she has no Facebook presence at all. And I was like, this is so automatable because of the nature of what it's doing. It's posting a picture of a sign, describing what's in the sign and then saying like, happy birthday or something like that. So I built in an afternoon, I built an iOS app for her that does exactly that. It just allows her to upload as many pictures as she wants. And then it uses the Gemini, uses Gemini to look at the images and generate captions. And then she can tweak the captions and then all of that is saved. And so each time the model is called, it's like, oh, by the way, these are the last 10 captions that the user accepted. So they like it when it looks like this. Right. And so it's, it's mixing it up. And so I did that inside of an afternoon. What was crazy about that is I had tried to build that app like a year ago at night. I was spending like two or three hours a night using AI. And I got like 80% of the way there and could not get it over the finish line. ([Time 0:10:16](https://share.snipd.com/snip/5781bc76-cdb9-48d4-8ea5-854e78ff310a)) - **Building Practical Personal Software For A Small Business** - Burke automated his wife's small-business workflow by building an iOS app that generates social captions and a routing app; both shipped quickly and replaced paid services. - The business runs on a Raspberry Pi and the AI-generated captions learned from the user's accepted examples to iterate tone. Transcript: Burke Holland And my wife owns a small business. Um, people who are listening to this pot to the pod may have seen this before, but like, if you're in the U S you may have driven by houses and there's like a big happy birthday sign in the Front yard. It's like different letters. Have you seen this? You may have seen this. Yeah. Yeah. Adam Stacoviak I'm in Texas. So we do things like that in Texas. We have a big front. Yeah. Burke Holland So it's a franchise, and she owns that franchise for our city in Tennessee. And so she has a lot of – What's it called? It's called Card My Yard is the name of the company. Oh, that's right, yeah. Yeah. Yeah, cardmyyard.com. Go order a sign for your – There you go. Do it up. For your kid's birthday or your dog's birthday. It doesn't matter. Okay. So she has, I build all these little apps for her. And these little like, her entire business runs on a Raspberry Pi. No way. Yeah. And it does things like, she gets orders in her email and it prints them out for her and stuff like that. But one of the things that they wanted people to do that the franchise wants people to do is to like get on Facebook and post pictures of the signs that people have put out in yards along With a caption. And like running a page for a small business is inherently marketing slop as it were, but it's like, it's a pain. People don't want to do it and she doesn't want to do it. Right. So she has no Facebook presence at all. And I was like, this is so automatable because of the nature of what it's doing. It's posting a picture of a sign, describing what's in the sign and then saying like, happy birthday or something like that. So I built in an afternoon, I built an iOS app for her that does exactly that. It just allows her to upload as many pictures as she wants. And then it uses the Gemini, uses Gemini to look at the images and generate captions. And then she can tweak the captions and then all of that is saved. And so each time the model is called, it's like, oh, by the way, these are the last 10 captions that the user accepted. So they like it when it looks like this. Right. And so it's, it's mixing it up. And so I did that inside of an afternoon. What was crazy about that is I had tried to build that app like a year ago at night. I was spending like two or three hours a night using AI. And I got like 80% of the way there and could not get it over the finish line. ([Time 0:12:01](https://share.snipd.com/snip/72070674-528e-452a-b8db-3c30266490b9)) - **Coding Is Easier; System Design Remains The Hard Part** - AI shifts the hardest work from writing code to architecture, security, deployment, and conceptual decisions; coding itself becomes easier. - Junior devs still must learn concepts like APIs, ETL, and systems thinking because agents won't replace architectural judgment. Transcript: Burke Holland I think that you still need to be a really high-performing developer to be able to actually ship. And the reason for that is that you need to understand a lot about architecture and security and how an app actually gets to production. Coding something is easy. Code has never been the hard part. That's always been the easy part, honestly, for being honest. It's getting that thing into production that's hard and that's still the case. And so this is why you see so many people talking about what they vibe coded and yet none of that stuff is in production because that's hard. So I think that you still need to know a substantial amount about what you're doing. What I don't have an answer to is like, so people who are junior, then the question is like, well, do you still need to learn how to code? I would think that you do because you can't. These concepts build on each other right you can't how do you understand what a rest api is if you don't understand how to call that rest api or how it's consumed underneath and the only Way to really understand all these levels of abstraction but but the second that that comes out of my mouth i'm like yeah but at some point you had to understand what the compiler was doing And you don't anymore. I don't know what the compiler does, nor do I care. So as you said, like, we don't know where all this is going, but I don't think that like, you couldn't bring in a non-technical person in and just drop them inside of an AI tool and be like, Do it unless it was like V0 or something for Vercel where it's really abstracting. But to actually ship a piece of software, no, you still need to be a dev. I think that for your high performing devs, I actually wonder about this because the thing is like in the past, you're really high performing devs or people that could do things like Build very complex algorithms to solve speed problems. So, right, performance issues, et cetera. But AI is really good at that, right? Like it knows all the algorithms. It can just destroy leak code. So I'm unsure at this point, if that's really the case, other than that, you need to know what to ask for a lot of times. Like, you may need to know that you need a binary search at this point, right? And to be able to direct the AI and say, why don't we do a binary search here and see if that's faster? You don't have to implement, you don't have to know what that is, how that's implemented. ([Time 0:18:50](https://share.snipd.com/snip/abeccf4a-ab5b-4750-a7a7-18f7fc7e022b)) - **Agents Accelerate Learning Of Systems Concepts** - Using AI accelerates conceptual learning: developers learn protocols and patterns (ETL, Unix sockets, gRPC) by iterating with agents instead of manual research. - This expands people's capability quickly, flattening barriers across languages and stacks. Transcript: Burke Holland That's a really good point. Somebody had made this point. I think I've made this point before that like, my fear was, oh, we're all going to stop coding and just prompt AIs and we're all going to get dumber because no one is going to know what they're Doing anymore. Because if you don't use it, you lose it kind of thing. But that's not what's not what's happening. What's actually happening is that, like you said, I'm learning all sorts of concepts that I never would have learned before. The world has expanded versus contract. Correct. That's exactly right. So instead of being overly focused on like, how do I compose this function? It's like, who cares? Do you know what a Unix socket is and how cool that is and how that works in Go? I didn't last week, but I do now. All right. That's a wild. And so I think that like as the people who are using AI are getting this knock on effect benefit that no one really realizes where they're just learning at an accelerated rate. ([Time 0:23:10](https://share.snipd.com/snip/3c200a4a-f6ad-440b-af88-1900db0155ee)) - **Classify Task Difficulty To Choose The Right Agent Workflow** - When designing agent workflows, decide task difficulty up front (easy, medium, hard) to pick appropriate agent behaviors and checks. - Hard tasks should trigger planning, sub-agents, model selection, and stronger verification steps. Transcript: Burke Holland I mean, part of what I do is try to figure out the workflows that developers actually need to be successful. And there's sort of two different ways to do this. When you're teaching someone, you really want to start like greenfield and simple. But in the real world, it isn't like that. It's more complex. So let's talk about like the first one first and we'll talk about the second. So for my Greenfield projects, I've experimented a lot with composing workflows that allow me to get from, okay, I have nothing and I want to get to this end point. What are the things that I need to do along the way and how involved do I need to be? Because you have all these different, you have the agentic tool itself, whether that's VS Code or the Copilot CLI. And then you have, or others, sorry, right? Cursor, Cloud Code, there's many others. And then you have like MCP servers, which is how your agent can talk to GitHub or how it can talk to Figma or, and then you have skills, which have serious overlap with MCP servers. And then you have instruction files and you have custom agents, which is you basically just overriding the agent behavior and defining workflows. And it's all very confusing for people because, because they don't know, they don't know what they're supposed to do with these things. And, and the honest truth is that neither do the people who are building them. Like, we don't know what the future looks like. ([Time 0:24:36](https://share.snipd.com/snip/afba83eb-b7a7-4a1f-9468-70783085b6af)) - **Plan With The Agent Before Asking It To Code** - Start projects with AI-driven planning modes so the agent surfaces missing requirements and options before coding. - Use iterative plan loops (agent asks clarifying questions) to reach a robust MVP one-shot then switch to autopilot for refinement. Transcript: Burke Holland So currently, my workflow is I'm, I'm now fully into the copilot CLI using it within VS code, but I start off with a plan where I basically talk to the CLI and I tell it kind of what I'm, what I want. And I'm in plan mode. And these plan modes have gotten really, really good. And people are like, I don't need a plan, but it's not about the plan. That's not what you need. What you need is the agent to think through with you all of the things that you just forgot to tell it that it needed to include in this initial prompt, because prompting is super hard. We did this idea of like prompt engineering, which is just prompt engineering is just giving the model the answer that you want. If you give it enough context, it will get the answer right and give you right. If you were to spell out exactly what to do, it could literally build Minecraft on its own, but you can't do that. You're that's not, you can't put the prompt to be, You can't do it. So in plan mode, you get this iteration where the agent's like, okay, cool. Here's, here's a, did you think about this? Here's what I recommend, but here's three other options, or you can do a fourth one. And you go through about like four or five, six loops of this. And I'm using Opus 4, 6 for this. It's incredibly good. And then once you get to a point where you're like, yeah, that's pretty good. Of course, you're not going to one shot an operating system, but for a lot of us, we can get to sort of like an MVP and a one shot. So you can like one shot to MVP and then iterate from there. So in your MVP, and then in copilot, we have something called autopilot and autopilots like a Ralph loop. You're familiar with Ralph loops. Yes. Where the agent is just orchestrator. Adam Stacoviak It's a rust project. Uh, it's really cool. Yeah. Burke Holland So similar, similar concept where it will, uh, basically feed the agent's output back to itself. It's just running the agent in a loop and it does this until the agent has some amount of confidence. There's this idea of like, if you tell agents, like do this until it's done, they're like, okay, I did it. It's done. You're like, it's not done. It isn't done. But if you tell them do this until you have 95% confidence that it's done, then they will run and run and run and run and run building up a confidence level. ([Time 0:26:19](https://share.snipd.com/snip/bd9bc7ea-027a-42d1-9655-9ce428b5d8b7)) - **Automate Iteration With Autopilot And Subagents** - Use autopilot loops with confidence thresholds (e.g., run until 95% confident) to let agents iterate without false early stops. - Orchestrate sub-agents and different models (Opus, Gemini, GPT) for design, refactor, and verification to scale complex tasks. Transcript: Burke Holland So you can like one shot to MVP and then iterate from there. So in your MVP, and then in copilot, we have something called autopilot and autopilots like a Ralph loop. You're familiar with Ralph loops. Yes. Where the agent is just orchestrator. Adam Stacoviak It's a rust project. Uh, it's really cool. Yeah. Burke Holland So similar, similar concept where it will, uh, basically feed the agent's output back to itself. It's just running the agent in a loop and it does this until the agent has some amount of confidence. There's this idea of like, if you tell agents, like do this until it's done, they're like, okay, I did it. It's done. You're like, it's not done. It isn't done. But if you tell them do this until you have 95% confidence that it's done, then they will run and run and run and run and run building up a confidence level. So So a lot of that stuff is abstracted into the tool. But then I'm also like using a custom mode, a custom agent in that that I call Anvil. And what Anvil does is it, it does two major things. The first thing that it does is it decides like, is this a, is this a easy task, a medium task or a hard task? Because you want different workflows from the AI, depending on what those are. Those are not, it's not the same thing. If you ask the agent a simple question, that's an easy task. It should just answer that. But if you give it a hard task where you need it to build something, then it really needs to plan. It needs to use sub agents. It needs to farm work out and get a copilot. You can use different models. So like you can use Gemini 3, 5.3 Codex, and Opus 4.6 all in the same run. And you can use as many of them as you want to. Okay, so now we're in an orchestration scenario. 4.6, if we're going to do a massive refactor, it might farm out the design work to Gemini, it might farm out the refactoring to 5.3. You might have 26 sub-agents running. You need to review that code. And so this is my workflow now is to try to automate all of that. So I'm planning and then I'm autopilot with my custom agent, which out the other side needs to give me something that's verifiable. Like, how do I know that you think that this worked? What is the evidence that you have that this worked? ([Time 0:27:44](https://share.snipd.com/snip/5ff63f6b-cc0f-481a-9100-50112e29faaa)) - **Run Agents Like A Team With Regular Reviews** - Expect to run agent work like a team: hold frequent reviews or standups with your agents to review progress and steer priorities. - Treat agents as teammates that need guidance, verification, and tradeoff decisions to ship production software. Transcript: Adam Stacoviak Here's what I really care about it. Here's all the things I don't want you to do. Here's the things I do want you to do. And I'm literally going to bed. I just told him what I'm doing. Came back the next morning and ledger's born. Yeah. Burke Holland I've thought about this a lot myself where it's like, what you really want is for the agent to just work all nights, like you said, and then build the thing that I want. But the problem is it doesn't know what you want. And you cannot get out everything because you don't know. And that's not how software is built anyway. In no world are all the requirements and constraints to find up front. What happens instead is that you think you know. You discover it as you go. That's right. And it changes and you make tradeoffs and the agent has to do the same thing. And so I don't know what the answer to this is other than whatever product you and I are replacing currently probably has a team of devs and an engineering lead. And every Monday morning they get together in a standup and maybe they have a standup every day where they go over all of the work that's in the backlog, what the status is on each one of Those things. And I almost wonder if that isn't the future, right? Because you can't just hand this work off and then just walk away and expect that the agent is just magically going to know exactly what you want. It just doesn't. In order to really build anything of size, you'd almost have to have a daily standup with your agent where you go over all of the things that it's working on and review it. And that's tedious. That's actually the part no one likes, right? Like Chad, if you're listening to this pod, like if you're in a daily standup, you're like, that's my least favorite part of the day. I mean, I'm with you, but there are reasons why we have processes and that's so that we can actually build things that work. I did something similar where I have a Ralph loop. I have co-pilot CLI running in a loop and it's building a multiplayer game where you just, uh, yeah, you just LARP as a baby bird. And you're hatched from an egg and you fly around. And I'm letting it. ([Time 0:37:25](https://share.snipd.com/snip/6e68e4e3-4e06-4854-8d92-61a5efb54df4)) - **AI Is Turning More People Into Builders Not Replacing Experts** - AI may broaden who builds software: more nontraditional roles will be able to produce working apps, blurring job titles between PMs, advocates, and engineers. - This doesn't replace expert builders; it democratizes contribution and increases overall software creation. Transcript: Burke Holland Somebody said the age of personal software is here and they're a hundred percent right. And I think that's what most of us are doing right now. And so for people who are listening to this, they're like skeptical. Like I can't stress enough. You can build anything. And if you're listening, thinking to yourself, no, you can't. I've seen what AI can do. The problem is that we've been trying, we've been on this sort of like hype grind for so long that we sold people like six months ago, like a year ago, we were like, you can build software With AI and you could, but it was tedious. And then a year later, we're telling people like, no, no, no, no, no, you can build software with AI. And they're like, fool me once, right? You're not going to fool me again. But I don't know what to tell you besides just go, y'all, just go do it. Like ever wanted to build a AAA game? Just go do it just go do it see if i'm wrong test me and see if you can't get to something that's functional inside of an hour yeah and you can you really can now these and really thanks to open Adam Stacoviak Source i mean we were and maybe you could speak to this given github and Copilot. There was a lot of outrage on the initial release of GitHub Copilot in that, oh, GitHub released this thing and it trained on all of our code. Damn you, GitHub, and all this stuff. But then open source is still here and still – I would say open source is thriving more than ever. There's this massive swing back to just simply open source and free and open tooling. I mean, Open Next is the example we just led with there. I mean, it's a good thing. Burke Holland Yeah, because it's so cheap to create software now that it's nothing for you to just create something and then make it available to everyone. Adam Stacoviak Yeah, I think – Until it's not. Well. Right. Not the cheap side of it, but until it's not worth sharing, you know, where you want to protect your moat or build a moat. Burke Holland A hundred percent. But I do think that there are emerging problems that we don't really have answers for. So like the VS Code is probably one of the biggest open source projects. And I think last year, every year when we get to December, they do like a triage of the open issues. They spend December just trying to do spring cleaning, essentially get the repo down to a manageable number and do clean up the backlog. And it was like a year ago, it was 8,000 issues. This year when we went into December, they were like, okay, how many issues are we working through? It was 15, issues wow like this is not we have to figure out the process and the team is actually doing this right so the VS Code team doesn't just build VS Code they're also trying to figure Out the AI processes for their own workflows that you use to create a tool which has to be reliable it has to be fast it has to work. It cannot go down, right? We can't just ship a release where people are like, the editor doesn't work anymore. We can't do that. Catastrophic. Right. Exactly. And so it's a, like you said, there's different kinds of software, but when we're talking about what we build for ourselves, that's one thing. But these teams, and I was just in Redmond talking with one of the engineers, Michael Lively on the team, who's just, that's what he's doing is trying to figure out what are the workflows And putting them together so that the team can actually ship software with AI and have it remain at the quality bar that they are going to insist that it is. That it just has to be a non-negotiable for people. Adam Stacoviak Yeah. Burke Holland Yeah, for sure. We should talk about your point about are developers going to be replaced. We should talk about that because I have thoughts here. Adam Stacoviak I know you do. Let's do it. So. Burke Holland What I see right now is this sentiment just keeps getting thrown out, but it isn't actually happening. Like AI is taking jobs, but not in the sense that it's doing them. It's taking them in the sense that like, if you're in tech, that is money that's being diverted to GPUs. Everybody's investing in GPUs and not headcount. And if you're in the enterprise, I think a lot of places are using it as an excuse to downsize, right? Like it's just being, it's a convenient scapegoat, right? But it isn't taking the job in the sense that like, oh, we don't need you because now AI does what you do. We're not there. Models are extremely capable. And so I'm pessimistic that that's actually going to be the case. And I actually wonder if we don't end up with more developers. But then you start to ask, well, what is a developer exactly? And we even have looked at this at Microsoft. Like, am I an advocate? Am I a PM? Am I an engineer? Or am I all of those things? Because I can now, if there's a feature that's missing in VS Code, I can open an issue or I could open a PR with three different implementations of that issue done with the AI. And then we can either approve and say, yes, we're going to do this. Or we can just be like, nope, we're not actually going to add this at this time. Close it out. What do I care? It's a 15 minute job for me to create three different iterations. So it's like this age of abundance where I can contribute in a place where I before didn't have the time to, I don't have time to sit down and learn the LLM API in the tool. I just don't, I have other things to do, but now I can actually, I can contribute. So am I an engineer? Am I an advocate? Am I a PM? And I think we're going to see more of this where like, what is it exactly that the job title is, right? ([Time 0:43:44](https://share.snipd.com/snip/18485da9-fd63-487b-b239-71445e149909)) - **Craftsmanship Will Be Rare Luxury While AI Builds The Commons** - Handcrafted, 'cathedral' software may become rarer and more expensive, but the craft will persist among a smaller group who command a premium. - Most users will choose cheaper, AI-assisted solutions; artisanship becomes a niche luxury. Transcript: Burke Holland You're doing to pull that off. If I was going to take the other side of the argument on what I just said, it would be that the reason why we recoil at this is because to us, to those of us that have been writing code for decades, It's a craft. It is a craft. You lamented about this in your post. I did. And when we had this big influx of people, and even my kids when they were in high school, they're in college now, but it was this big push that everybody should learn to code. And literally everybody was in these coding classes because they wanted to make a ton of money. That's not a bad reason to want to code, but they had zero interest in the craft. And that's why so few of them just made it out the end of the pipeline is because it's hard. It's hard and it isn't fun and it's tedious and you really have to enjoy the craft. And it's like AI is never going to write code that's up to our standards because it's not a craftsman. We're craftsmen. AI is not. And so I do worry that we're building essentially like strip malls and track homes, right? Whereas like, if you go back in history in my town here, they were built, the homes that they built were just beautiful. And if you fast forward to today, they're nothing, right? Even my house is just a brick square. And that's fine. I don't want to be ungrateful. But at the same time, can we even build? Can we build cathedrals, Adam, anymore? Does anyone even know how? I don't know that we do. And so I worry that we're not, we're going to forget how to build cathedrals. This should be the next blog post. ([Time 0:54:08](https://share.snipd.com/snip/5a668ee5-673c-48ed-aa80-1fd887cc6ba2)) - **Incumbents Benefit From Model Familiarity Even As Reimplementation Gets Cheaper** - Established incumbents gain resistance because models already know and favor existing tools (NPM, Firebase), making immediate replacement hard despite easy reimplementation. - New alternatives must supply compatibility and teach agents quickly to gain traction. Transcript: Burke Holland But that is inherently super dangerous. Open Claw is one of the coolest things I've seen. What a great project, right? And you can lock it down. But you kind of have to do that. You have to lock down, say, well, only this Telegram usergram user. And then you have to test and make sure that that's actually the case. Did that actually just happen? Or did it just tell you that it did it? And, but I think that when it comes to things like NPM, one of the things that NPM has going for it is that these models are, they know so much about it and there's, they know how to use it. They know what these modules look like. There's all this history. It's in the training data, right? The second that you invent something else, well, now the model doesn't know what that is. And you can mitigate that by providing context and maybe future iterations will know, But it's like models are so good at using these established tools. For instance, like I know I work at Microsoft, but if you want to have some fun chat, you can use Firebase. Models know everything there is to know about the Firebase CLI. So you can literally log in on your machine and then tell the agent to build something, tell it it can stand up its own Firebase services, which are free, right? You have to physically enable billing here. And it can do everything. It knows how to use the Firebase CLI from front to back. That makes it really hard for you to supplant that with something else, because people choose the path of least resistance. That's my immediate reaction to that is that it's just too easy. It's already there. Models already know how to use it. ([Time 1:00:51](https://share.snipd.com/snip/a0fb691d-732a-4b01-96f6-fb728b3dff39)) - **Combine Models For Planning And Senior-Level Coding** - Multiple models have complementary strengths: use Opus 4.x for planning/interaction and delegate tougher coding tasks to GPT-5.3 or Gemini for higher-quality implementations. - Orchestrating models gives you conversational guidance plus senior-dev-level code generation. Transcript: Burke Holland And we should point out that opus 4.5 kind of like to me kicked off started something but it is not the only model right there we now we have 4.6 there's 4.6 fast there's 4.6 with a one million Context window five three is an outstanding model context window yeah five three is unbelievably good uh gemini is a great model right they just released three one great model right So these model companies are really going at it trying to deliver the best coding experience and then developers listen we're the ones who win right we're getting all the best tools. We're getting all the best models. Now's the time. So I agree with you. I've never built so much software in my life and I'll have an idea. I'll be like, oh, I wonder if I can do that. And then I'll just go do it. ([Time 1:31:23](https://share.snipd.com/snip/3532fa40-9c04-4ebb-8c2a-ebba052a2591)) - **Experiment Early With Small Paid AI Plans** - Don’t delay adopting AI tools; start with the smallest paid tier to experiment and scale usage as you learn. - Graduate usage limits intentionally so you get feedback from real constraints while iterating rapidly. Transcript: Burke Holland Well said. Adam Stacoviak Well said. I'm personally getting some really good enjoyment from Codex. GPT 5.3 has been really cool. I feel like when I play with that thing, it spits things back to me. I'm like, can you explain it to me like I'm five? Because I don't even know what you just said. You're just so next level in how it responds. I'm like, you think I have the context here, but I'm not following you because you're way smarter than I am. Burke Holland Can you please just dump it down a little bit? Described five three is like a really really senior dev and like opus four six is like your best friend yes and so they're they're different in the way they interact and so like i said i'll Use i use four six i interface with four six and it just delegates work to five three because i actually don't want to i with 5.3. I know that it's a really good dev, so I want it writing code, but I want to talk to 4.6 because it's always like, oh, yeah, you know what? You're totally right. I'd miss that. You're so smart. We should talk. That's the gaslighting. It's so nice to me. It's like, yeah, you're so smart. Yeah, it makes me feel good about myself. Adam Stacoviak I'm with you on that. I've been actually doing that exact step, which is I've used Opus 4.6 to do planning and then delegated to 5.3, GPT 5.3, for those exact reasons. Because I'm like, you know, I thought I was smart, but now I'm talking to you and I feel like I'm just not that smart anymore. Or it's just not as nice. It's like, Adam, can you just pick it up here? I'm like, I want to be 10 steps down the road here and you're holding me back here yeah yeah which is uh which is part of the battle here what else you want to say burke i know that we got lots To say you got some cool stuff we're working on the github co-pilot cli we didn't even mention that really i think it's a beautiful thing uh great work on that front there i mean there's A the world's the oyster for the next coding cli to kind of keep iterating. You got open code out there. You got Claude, you've got Codex, you've got get a copilot CLI. You got all the things out there. What do you want to say in closing? Yeah. Burke Holland I would just like to say, you know, for me, because I've been working on the copilot CLI, that's exciting. And it just went GA today. And that GA is just our way of saying like, it's ready for production use because we have so many enterprise customers and they won't touch anything until. Yeah, just today. Wow. And they can't use anything until that thing is out of preview. It has to be stable, right? They have to know that it's supportable and usable. And so today you can start using that. Super excited about the stuff that is being done to make AI even better so that you don't have to compose these like hacky workflows or make Ralph loops, whatever the heck that is, or use A gas town. Like there's all these crazy ideas and terminology being thrown around. And the truth is that like, you don't have to use any of that stuff. Just start working with the agent and see what works for you and create your own. You're smart. You'll figure out the workflows. You can make the agent do anything. Don't, don't, I know it's easy to get overwhelmed. Trust me. I know. Yeah. Very easy. But just, just use it and figure out the workflows that make you most productive. And you're already, if that's all you do, you're already in like the top 5% of folks at the moment. Yeah. Adam Stacoviak I would say don't delay. Really, really don't delay. If your pocketbook can handle it, get the smallest version of things. If it can go further, go further. And I would do that based upon graduated usage, of course. ([Time 1:32:29](https://share.snipd.com/snip/776a7a06-ddcf-4c02-bd12-477696827543)) - **Choose A Practical Agent Workflow And Iterate It** - Pick workflows that fit you: try Copilot CLI, VS Code agents, or CLIs that let you plan, run autopilot loops, and verify outputs. - Iterate on personal patterns until you reach a productivity groove; early adopters are already in the top percentiles. Transcript: Burke Holland I would just like to say, you know, for me, because I've been working on the copilot CLI, that's exciting. And it just went GA today. And that GA is just our way of saying like, it's ready for production use because we have so many enterprise customers and they won't touch anything until. Yeah, just today. Wow. And they can't use anything until that thing is out of preview. It has to be stable, right? They have to know that it's supportable and usable. And so today you can start using that. Super excited about the stuff that is being done to make AI even better so that you don't have to compose these like hacky workflows or make Ralph loops, whatever the heck that is, or use A gas town. Like there's all these crazy ideas and terminology being thrown around. And the truth is that like, you don't have to use any of that stuff. Just start working with the agent and see what works for you and create your own. You're smart. You'll figure out the workflows. You can make the agent do anything. Don't, don't, I know it's easy to get overwhelmed. Trust me. I know. Yeah. Very easy. But just, just use it and figure out the workflows that make you most productive. And you're already, if that's all you do, you're already in like the top 5% of folks at the moment. ([Time 1:34:22](https://share.snipd.com/snip/96e9a324-0aff-4d49-b9cd-2529fecedb6d))