Warp CEO Zach Lloyd on Boosting Developer Productivity with Generative AI in the CLI

Hey guys, thanks with Throw, we're here with Zach Lloyd, who's the CEO of Warp, and we're having a little chat about something called Warp Code, which is a whole new way of thinking about CLI for application development in the age of ai. Zach, welcome to the show. Thanks for having me.

Excited to be here. All right. Well, I'm not sure everybody knows exactly what Warp is or what Warp Code does, but maybe walk us through the fundamentals here and why do we need a different way of thinking about the CLI?

Yep. So the, I'll start with just like the basics of Warp itself. So Warp, um, started, uh, three or four years ago with the idea of, uh, improving just like the fundamental terminal interface.

The, um, you know, the kind of genesis to the idea was I've been an engineer for a really long time. I've always worked in the, in the terminal, never been a terminal power user, but I've always worked with people who were, um, really good at using the command line and wanted to build something that made that power more accessible to other developers. So we spent the first couple years of warp just trying to make the terminal more usable.

Um, when ai, uh, started becoming more powerful, uh, even before sort of chat GPT, uh, we started building features into Warp that would allow developers to, uh, sort of translate natural language into terminal commands. So if you're like, oh, I don't know how to, um, you know, find files across my directories or do certain advanced things in gi, you could, uh, sort of ask or in English how to do that, and it would give you the terminal command. Uh, as LLMs became more powerful, we realized like the terminal interface itself is actually a great interface for interacting with ai.

And so instead of just using like terminal commands to drive your computer, uh, you can use, the biggest thing that's different in work versus another terminal is that you can just use natural language directly to ask your computer questions or ask your computer to do things. So, for instance, you could be like, set up a new project for me with React and TypeScript and Warp will do it for you. The, um, as like, um, the models become more powerful, we realize like people want that sort of interaction, not just for doing things that are traditionally terminal tasks, but just doing coding.

Um, which is probably like the number one developer activity. And so the most recent thing that we've launched in Warp is a way where you can do a gen coding. So you can, you know, simply tell your computer what you want it to build, what feature you want it to build, what bug you want it to fix, uh, in natural language.

And it looks a lot like a terminal interface, but it actually can do things that are more typically in a code editor. So like it can produce diffs, it will show you a, uh, way of like reviewing your code. So it shows you the active Diff, um, it, uh, basically makes the coding experience where you're, if you're starting a coding task with a prompt much better.

How does that differ from all the other AI coding tools that are out there? Are they kind of trying to wrap some sort of graphical environment around it and most developers don't seem to enjoy? Or what's the difference there?

Yeah, so there's really, um, the way I think of it, there's two other kind of buckets of of competitor products that have different approaches. So one is like the kind of AI enabled IDE. And so this is something maybe like Cursor or uh, GitHub co-pilot where the fundamental interface, uh, is still like a code editor.

So you're opening up files, you're looking at code, you're handwriting them, and kind of the best feature in those apps, in my opinion is like AI driven, auto complete. So you, you type in like you see the ghost text of like the computer, uh, of like, of the AI suggesting what, uh, should come next. Those tools also do have like agents in them, but they tend to be in a sort of chat panel.

And, um, you know, it's, it's a less natural way of working with agents than just having like the whole interface dedicated to the agent. The, um, the other thing that's really, I think, more usable in Warp compared to those types of tools is like, um, you know, you can do stuff, uh, across multiple projects at once, whereas the IDs tends to be very like one project at a time workflow, and you can do things across the whole software development life cycle. So Warp is really good for setting up new projects for coding on them, but then for also like deploying them, debugging them in production for interfacing with like your cloud services.

So the terminal is like kind of like one level deeper in the stack. Uh, and so that's makes it, I think, a better option than the IDE based alternatives. The other set of tools that have caught on lately are things that are like, um, cloud Code or Gemini CLI that are like, they're not terminals, they're apps that run within the terminal, so they're like CLI apps.

And so those are, um, it's a very, it's a similar way of working to Warp, but with a much more limited, and I think hard to use interface because you're limited to, uh, just like a purely text-based app. And so, you know, like you, you don't get editable diffs, you don't get the ability to review, uh, an agent's code in a code review pane. You don't get a file picker.

So there's, there's all these like kinda limited things around working in a purely tech-based environment that I think, um, you know, warp makes much easier and, and gives you a better developer experience with. Is this also likely to appeal to software engineers that are running DevOps processes? 'cause they typically work within that terminal as well.

And I think a lot of the AI coding tools are a little more tuned towards traditional developers without thinking through maybe what software engineers need. Yeah, I, I totally agree with that. So that's one area where Warp is like really strong is for DevOps and production for people who are more comfortable in general and the terminal who are doing tasks that are more terminal oriented tasks.

Um, I think the other tools also are more widely adopted by like Vibe coders, so, you know, people who don't even really necessarily know how to code. And we have some people like that who use Warp, which is cool. Um, but we're focusing more on pro developers who want to, you know, get to shippable AI generated code, which means like, you know, as a developer you really need to understand, uh, what code an agent is writing.

You need to be able to review it, you need to comprehend it and like be able to stand behind it as though you wrote it yourself. And so we're, we're not really focused on the use case of like someone who doesn't know how to code making a whole app in Warp. We're much more focused on the use case of a pro developer who wants to use, uh, AI to accelerate their daily workflow.

Mm-hmm. And what is the impact of AI so far that you've seen on terms of what it means for professional developers? 'cause to your point, I'll talk to them and they like the idea of it, but then they struggle when they have to debug an application 'cause they didn't write the code and they don't have a lot of understanding of what's happening in there.

And so then they get a little frustrated. Yeah, this is, this is a hundred percent the biggest problem with it right now. So if, like you said, if you look at the Stack Overflow, the latest Stack Overflow survey, um, the top two complaints, a, about AI generated code from pro developers are that the AI produces code that's hard to understand and has subtle bugs.

And then the second thing is like, it's really hard to debug and work with code that you didn't write yourself. And so, um, these are, I I think there's a couple approaches to like fixing this, um, and being productive with, uh, with agents. The first thing is like, I think it's like a, almost like a cultural thing or, or just like a how do you as a developer approach working with these tools if your approach is like you're just gonna like tell an agent to build a future and expect to get good chipp able code out of that.

I think that's not really where we are today, frankly. Uh, it might work for like small, for small things where it can sort of like, just like do it in one shot it, but more typically, uh, if you take that approach, uh, you're gonna get code you don't understand, that might be buggy that probably won't work, or maybe it'll work, but you won't understand why and it won't be mergeable. So what you have to do as a developer is change your approach when working with AI to first off, specify not just like what you want to build, but you have to specify how you want it built.

And so I think that means like, um, you know, working with an agent on a, uh, first of all, just like understanding the code before you even ask you to build anything. And so these agents are actually like awesome code explainers and like code explorers, like they can, they can help you get familiar with new code really, really quickly. Uh, and then the next step is like, once you understand the code, it's to be, you know, iterate on a plan, uh, or a spec, whatever you wanna call it, where you get really clear with how you want something to be built.

Uh, then you, you know, then where the agent can really save you time right now is in like, implementing that plan. Uh, and it, it, uh, you know, you don't get as many surprises if you do it that way. The third thing is that you should work in small chunks.

And this is just like good engineering practice in general. Like, this doesn't really have anything to do with agentic development, but like, if you try and do one huge thing with an agent, you're gonna have the same problem as if like you wrote, you know, one huge PR on your own, you'd never want, like some, you'd never wanna review a giant PR from someone else on your team. And so, like the way to do it with agents is, um, you know, specify a small piece of it, maybe you have it, like write a few functions or write the model layer or just write a, a part of the API verify that it's right, write tests for it and continue.

And so like you have to like kind of think of it as like you're guiding like a junior engineer to write code. Uh, and if you do that, I do think you can get really great, uh, productivity gains out of it. It's just like you have to have like a, a reasonable mindset around like, you know, how to work with it and not expect sort of magic.

Mm-hmm. The, um, and so the, you know, I guess one of the big things that we're focused on feature wise at Warp right now is like, how do you make that workflow really, really seamless where it's not just like fire and forget and let an agent just like go off and run and like do your thing, but like, how do you review its code as you go? How do you work with it to get to a good plan?

Um, and so those are the types of product features that, that you want. If you're a pro developer doing agentic development, How will this all evolve in your mind? Is each developer gonna have one AI agent that does a bunch of things?

Or will there be multiple AI agents that are trained to do various things? And if I'm on a team, well then we all have to coordinate our AI agents amongst ourselves, or maybe there'll be AI agents on the team that are trained to do specific things on behalf of the entire team. Yeah, so it's a great question.

So in the short term, I think it's totally fine to have like every engineer has their own agent set up because, um, it's like, it's very much the agent is in the inner loop, meaning like the agent is in like the part of software development before the engineer even shares their work. And so it's, in a similar way, it's fine for every engineer to have their own different IDE set up. I think it's okay right now for every engineer to like use their own agent to, to help them get to a, a pr, they want someone on their team to review.

I think as agents move to the cloud, which I think is starting to happen, um, it makes more sense to standardize. And so that, you know, like I think something that's gonna happen is like you're going to have agents writing code based off of system events, meaning like, oh, there's a crash that your crash reporting system has detected. Okay, an agent is gonna be listening to that and write a draft pr and that's not even involving like a developer per se in their personal setup.

So for something like that, I think it makes much more sense for people to standardize, but we're very, I think we're very, very early in that workflow, uh, and where we are today. It's like, it's much more like, can you get developers accustomed and comfortable working with the genic workflow in a way that suits, um, how they like to work. Mm-hmm.

What do you think will happen with professional developers? 'cause every time we turn around somebody's saying, well, this is great, we won't need professional developers. But as far as I can tell, somebody still has to at the very least, manage all the AI agents.

So how, you know, from your perspective, you know, what's reasonable expectations here? Yeah, I think, um, I don't think professional developers are going anywhere anytime soon. I think, um, if you're a professional developer, the smart thing to do is like invest in learning AI as another tool.

Like, I think that's, that's where we're at with it right now. It's like, it's a tool and if you learn how to use it well, you can become a way better developer. And so I don't think it's like a great strategy as developer to be like, I'm, I'm just gonna ignore this because it's like, yeah, I don't know, what's a good exam?

It's like, oh, I only wanna work in assemble, or I'm not gonna work with compiled languages. Like, that's like a self-defeating, self obsoleting approach. But I also don't think that there's like, there's not like real risk, like at the moment, uh, every company I know is hiring more pro developers.

Um, warp is hiring more pro developers. I think if you are, um, a pretty inexperienced developer, there is risk. Like, 'cause like, it's the same kind of risk that exists from like WordPress or website building platforms, displacing people whose primary skill was like building websites.

Like, like if your primary skill is like building something that an AI can totally build all of its own, I think you need to uplevel your skills. But for by and large, for people who are working at like, on either enterprise software or complicated consumer software for their job, I don't think there's like super big risk right this minute. Um, you know, the models change quickly, so I won't say like forever, but I don't like we're hiring engineers right now.

It's very competitive, so I don't, it's like it hasn't really changed. So what will be the ultimate outcome? And I ask this question because some folks say we're about to build more software in the next two years than we built in the last decade.

But in my mind it also just seems like, well, maybe we might build software a little bit faster, but maybe we'll just build better quality software with less stress. Yeah. So I think it depends like what kind of software you're talking about.

So there is this submergence of this new kind of software where never would've been economical to build software before. Um, but an, an agent can build it for you. And so this is almost like personal software.

Like I know people who are building like ad hoc things to manage like their wedding or whatever. And so it then, who are not developers, they'll go in, they'll use like a tool and they'll use something like they could use Warp, but they might use something like Lovable or Bolt or whatever, and they'll, they'll build an a disposable personal app that will only be used one time. And so that's like a whole new kind of software that never would've made sense to have been built before.

So there's, there's gonna be more of that in the like professional context. Um, I think it's more like what you said, like I think it's like a productivity multiplier if it's deployed correctly. Um, I, uh, you know, I don't, I don't, I have yet to see someone build from scratch.

Something that's like a truly, you know, hard to build app like Warp or Figma or Google Docs or like what, like it's just not a thing. It, like, it's, it's too hard and for the agents to do right now. Uh, and so I think what you'll see is like AI gets deployed as a tool.

It's, there's like a percent acceleration, which depending on the type of task you're doing might be, might be small or might be really big. Like it tends to be really big in circumstances where, um, you are doing something as a professional that's like, would it like zero to one, like in a language you don't know. Like there are all these cases where it can really accelerate, but for day-to-day development, I think it's like some percentage acceleration and we will, you know, help companies ship better software more quickly.

All right, folks, you heard it here. The way we build software is definitely changing, but the key thing about all that is, well, there's still humans doing it. Zach, thanks being chef.

Thank you for having me, Mike. This was great. All right.

And back to you guys in the studio.