Warp's CEO on What It Actually Looks Like to Build with Agents in 2026

0:00This is a bold prediction, but I actually think that like something like Warp or Cloud Coworker, that's going to be the interface that knowledge workers use for a lot of stuff. Rather than using all of these like individualized SaaS apps, it's just going to be a personal agent control plane, which is pretty cool.

0:20Matt Paige:Welcome to the Talking AI Podcast, where we talk AI with both experts in the field and early adopters. I'm your host, Matt Page, and we're here to demystify AI for you so you can get some value from it. Let's talk some AI. We have engineers on our team who haven't written a single line of code in three months. And a year ago, that would have gotten somebody fired, but now it might mean that they're the most productive person on your team. Agents and agentic approaches are completely changing what it even means to be an engineer, how we build, how we think, and honestly, how we define the role itself.

0:56Matt Paige:And I got the perfect guest to talk about this with me because Zach Lloyd is the former principal engineer at Google, former CTO at Selfmade and Time, and is now the founder and CEO of Warp, an agentic development environment built for this era. Welcome to Talking AIs. Thanks for having me. I'm excited to be here. Yeah, we may show some of this off. So for those just listening, you may want to catch this on YouTube. But Zach, it feels like we've crossed some threshold towards the end of last year. And maybe it was Opus 4.6 or Codex 5.3, maybe just the agentic tools maturing. But the actual act of building software just looks completely different now.

1:39Matt Paige:How would you describe that shift that we're experiencing in real time? Yeah, I would agree. And I think that the timing you have is right. I think if you were to look back, I don't know, a year ago, you would start to see AI really changing workflows. But it was primarily with things like autocomplete and you'd still be in your code editor looking at code. And you'd have these cool assistive features where, you know, AI would sort of complete your thought. In the middle of last year, it started changing to more like you just prompt your computer asking it to make changes for you. this was things like cloud code or warp has this ability as well and uh it started to change the workflow where you could start everything with a prompt but you probably wouldn't be able to finish it uh and then i would say towards the end of last year like your your timing is right the underlying models i think you picked the right two ones with opus 4 6 and codex 5 3 have gotten to the point where it's like you really don't need to write code you still need to like understand code as a developer.

2:49But if you're handwriting code, you're either kind of doing something wrong or it's like you're working on a really hard big code base. Like there's some circumstance where you're not doing like straight down the line engineering because I think for the most part, even on like pretty hard code bases, as long as you're guiding the agent, you can have it actually like write most of the code for you right now, which is wild. It's like sci-fi.

3:13Matt Paige:Yeah, it is crazy. And, you know, you'd see these people online that would like zero, you know, single shot, um, some app that exists in the world. And it was kind of a joke because it was just a front end. But today, like it's just a different experience. Like as I was playing around with warp before this episode, like it goes through the step of actually planning what you're going to build. Like it truly feels like the experience of working with another engineer. Like you're aligning on, okay, well, what do we want to build? What about these features? What about this tech stack before you even get started?

3:49Matt Paige:And then it just goes in like you, the human continue to be a feedback loop, but it's like the agents are a feedback loop onto themselves in a sense as well. But like, what's the comment? This is like the new emerging workflow where people are spinning up these multiple agents. And I think that bit is foreign still to a lot of people where you have agents working in parallel. Like, yeah. How does that work? How are engineers leveraging that? Like, what's this new kind of working model with building today? Yeah. So in today's world, the way that developers are working is, I would say, primarily with agents like on their laptop, like kind of local agents.

4:35You will do pretty much what you said. So the way that I work is like I will pick some task I want to accomplish. Maybe it's fixing a bug. Maybe it's building a small feature. I will start like a planning process with an agent. Sometimes you can just straight up ask the agent to do something. And if it's simple enough, the agent will just do it. I like to work with the agent to understand what it's going to do and how it's going to do it and make sure it's right. because there are still, if you don't do that step and you're working on a complicated code base and I'm mostly working on warp, which is like a million lines of rust, the agent will still often do things and they'll kind of seem right.

5:16But when you look under the hood and you go to send like a code review to a teammate, it'll be wrong. But like, you know, it used to be that they would like delete tests or like make them trivial. Like they'll do stupid things to try to get to the outcome that you've described. So I would spend time planning and then I'll launch the agent And then, you know, it's now the case that the agents can run 10, 20, 30 minutes, 40 minutes. Like if you start to get longer than that, I start to worry about the agent going in circles or losing context, which is still a major problem with these agents. But if you have something that's running for that long, you're probably going to want to go do another one and then maybe even another one.

5:55And so at any given time, my workflow these days is like working on with three or four agents. I'm like it's actually like very taxing on me as an engineer to like keep track of what they're all doing making sure they're doing sensible stuff it's taxing on my laptop uh like you know I'm like spinning the cpu and like you know I have a whole bunch of checkouts so it's like a weird new workflow where I'm trying to juggle agents is kind of where we're at today I think that's going to change this year I think it's going to evolve again but that's today's workflow and you mentioned

6:30Matt Paige:to that point around because people are doing this locally they're spinning up a bunch of agents and while one agent's working on something you have another one working on something else but i think that's the interesting point i hadn't really thought much about is just the taxing nature on your actual machine right and you see like the the hot thing now is open claw and stuff like that you go buy a bunch of magnetes but those are also very expensive and i think that's a nuance with Warp and then Oz with your building. It's almost like agents in the cloud, in a sense, versus having them always be on your local machine.

7:05Matt Paige:But what's, you mentioned that shift happening. Like, where do you see that happening? And what's the, what do you foresee coming? So I think six months from now, and it's, keep in mind, it's hard to predict because this stuff is just changing so quickly. But I think there's a very strong pull to move agentic development off of people's local computers there's a strong pull from like the individual perspective where you know once you're running three or four of these it's hard to it's hard to do other stuff like they literally are running the limits of your laptop are you going to buy a mac mini like you're not going to keep scaling your individual compute by buying more computers i think that's like a crazy model what you're going to do is be like how do i take advantage of the fact there's all this cloud compute in the world.

7:53And you're going to, as an individual, want to run agents doing like workflows on cloud machines. As a company, you're definitely going to want this because right now it's very hard to see what agents are doing when they're running locally. It's very hard to like make sure they're secure. It's hard to like compound the knowledge of an organization. If every time an agent does something that like trail of what it did just kind of disappears into the ether. And so so there's a big I think there's a bunch of pull and we're seeing this with what we just launched Oz for companies to sort of like pull these workflows off of laptops into the cloud.

8:37And then there's just some other functional reasons why you want these things in the cloud, which is like agents are going to be running off of triggers like timers or like system events like, hey, your server crashed. And that kind of has to, by its nature, run on some cloud machine. It just doesn't make sense for that to run locally when the trigger is something that's in the cloud. So my prediction is that, you know, through things like Oz, you know, these workflows will move into the cloud. I don't think they'll move entirely into the cloud. There's still like nice reasons to do things locally, but more and more of it's going to move into the cloud this year.

9:13And it's going to be like orchestrated and auditable. And like these things are going to be more like cloud team members than just, you know, you interacting with them locally.

9:25Matt Paige:Yeah. It does have that vibe. It's, I forget who said it, but it's almost equating to like, you're going from individual contributor to like an engineering manager. And now you're having to delegate these tasks out. And there's that element of like, you know, you got to let go in a sense in terms of like, okay, you know, you got to trust their judgment that they're going to do the right thing. You still oversee them and whatnot, but you can't always get to looking at and inspecting every line of code. which like i'd be curious like do we abstract away to where like engineers actually getting that nuance into the code base goes away to an extent or that always exists as we just scale and start exponentially faster it's an awesome question so yeah like i think the question is something like will today's code become the same will the same thing happen to like what happened with assembly language.

10:22So it used to be that developers, you know, would look at assembly language and like that level of abstraction is gone except for the most specialized people. I think something like that will happen with a lot of code, to be honest. You will, you'll just think in higher and higher levels of abstraction. You'll think in terms of the systems and the architecture and like the stack and the services that you're using like if you're doing development like that's like where i'm operating now a lot of the time when i'm working with agents and that stuff's not easy by the way that's like stuff that you learn through a long career being an engineer so it's going to be interesting to see how people change their approach to like learning how to how to build systems because that's like that's the fundamental skill now not the code writing um we're not there yet like everything that everything that an agent writes right now still gets like a human code review and i would recommend doing that just because there's so many ways that agents can do things incorrectly but i i don't know i think it's definitely where we're trending is that you're not looking at code which is crazy to say but Yeah, totally.

11:38Matt Paige:I think the context piece is critical too, but I want to hit this first. Like, yeah, the user interface itself, the UX and UI of it, like what's your thoughts on where this is going? How are you thinking about it at warp? Because you have everything from lovable, which is in bold, which are, you know, fully web-based. It's like this nice user interface. It's kind of abstracting and doing everything for you. You're not really getting into the code. You have Cloud Code, which is more silly, you know, command line terminal based. You have Codex, the app that they just came out with, which is like something in between.

12:17Matt Paige:What do you think is the future? What is Warp? And like, how are you thinking about like the design of how humans interact in this new way? Yeah, it's an awesome question. So historically, you know, Warp has been a modern terminal. um like that that's what the company we our first product was it was like can we modernize and improve the user experience around the terminal can we make it work for teams um we have you know my attitude on this these days is like warps in agentic development environment it's it's grown out of the terminal that's like a agentic development environment is interestingly enough that's a term that we coined i think that is catching on as like a category the way that I think about it is we want to build a tool that is most suited for the new workflow and the new workflow is around prompting and then increasingly it's going to be around sort of like having a control plane for what all your agents are doing whether you've directly prompted them or they're running off some system event whether they're the back end for some app they're an automation etc so I think you want like the interface in the future is going to look very much like you can see what all of your agents are doing you can see what contexts they have you can see who launched them you can see what prs and artifacts they've produced and then if you need to you can jump in and like you know get personal with them so you can like uh you know take over the code you can iterate with them and so i think there's like a sort of control plane interface that everyone's going to have and that's kind of like what we've been leaning into with oz which is actually decoupled from the warp terminal.

14:01And then I think there's a place where, you know, if you want to actually do a tight iteration loop, it's going to look more like, it's going to look like probably the, I don't think it's going to look like the IDE. I don't think you need the code editor. I think it's going to look mostly like a terminal, but with some features that are, you know, sort of built for the agentic workflow, like being able to quickly look at diffs or look at your file tree and stuff like that. So I think that's what Warp is tending towards right now is like the, like a really good, uh, interface for doing this tight loop prompting with agents.

14:35Yeah.

14:36Matt Paige:I saw something, it was today, it was Cloud Code. They have something called remote control where you can basically interface with the agent on your mobile phone while it's still working in your local environment. And like, I think Spotify engineers, they talked about something similar where they're like interacting on the train. Um, yeah. Yeah. I'm wondering if there will be this other element too, where it's, you know, this, this form factor that is, and that probably won't be the typical thing where somebody's only doing this, but you could, right. You have your agents kind of working and you're conversating with them, um, through different form factors.

15:13Matt Paige:And I think voice is the other element too. Like, I don't know about you, but my default now is, uh, whisper flow or whatever, you know, that's, I've kind of just normalized around that. It was tough for me to make that transition, but I think that's, what's your thoughts on the other elements of like inputs and things like that? Cause it's becoming very conversational, I feel like in some ways. A hundred percent. So we, Whisperflow is built into Warp, which is cool. So you can just hit the microphone button or once you do that, once you can hold down a hot key and that's how I prompt. I just find it so much more efficient than typing long prompts.

15:50So I'm 100 % with you on voice. And then Warp has that same feature that Cloud Code has, and we've had it for a long time. It's actually annoying to me to see Cloud Code launching that. The PM on Cloud Code used to work at Warp. Oh, you're kidding. No, I'm serious. He was kind of poached by Anthropic. And so now they're launching it. But we've had that for a long time where you can take any local session in Warp, grab a link to it and then open it up uh open it up on your phone you can do that from a local session or i think what's really useful and actually a better feature is you can do that from uh from cloud session so anything from oz uh you can you can grab a link to it and it's all behind access control too so it's like you can share with your team you can have multiple people from your team observing agents work and i do think that's part of the future it's like you can start you can launch agents on your phone like you could be playing golf and having agents running um it's uh they're just going to become more like these things that are constantly working which is again it's like a crazy sci-fi thing but that is really what's going to happen so okay you just triggered me to kind of go down another rabbit hole but the competitive landscape

17:05Matt Paige:right now like you're literally in the space competing with anthropic open ai um the other hot vibe coding tools lovable um bolt all of those like what does that competition look like because y 'all been around for a decent amount of time like any insight into what it's like Like literally like competing with these giants? Yeah. So, I mean, we're doing well. We have our own sort of angle on things, which is, it's really the differentiating thing is like the tight, tight terminal integration that we have. Just to like lay out the competitive landscape a bit. So we, I wouldn't say we compete much with like Lovable and Bolt.

17:51So we are more of a tool. Yeah, we're more of a tool for like professional engineers. We also get a lot of professional, like, just knowledge workers, which is interesting, but it's more professional use cases. So we do get, like, PMs and data, like, people working on data teams and, like, but it's almost all business use cases. We do compete with, like, Cursor, Cloud Code, Codex. By far the hardest thing for us there is not just that they're, like, bigger brands, which they are, but specifically for like the model providers who also are in the coding app space they like they just can offer a much less expensive product and so you know to get people to pay for warp you know we really have to prove that the product is more useful and for a lot of people it is like people still pay us we're growing very quickly but like the the sort of like amount of pure ai usage that you can get from like your Claude Max subscription or your Codex, your ChatGPT subscription is higher than what you can get with Warp.

19:00And so the way that we are positioning more and more is to sort of be complementary to these tools. So you can run them within Warp. Like a lot of people will run Claude code or Codex within Warp. And that's still like, that's a cool place to be for us. We're in the sort of value chain of like people building stuff with AI. we are uh you know for orchestration in particular we are um uh going to be supporting like orchestrating with any of these other tools as well and so you know going more towards the enterprise orchestration market i think is is a better is a better spot for us to be but it's it's a super competitive space um the other like i will say two other competitive dynamics i think are like good for us in the long run is like one i think it's just a gigantic market like Totally.

19:49Ride that natural wave. Like in like the, it's in its expanding market super quickly. Like who can build with warp or cloud code is like expanding rapidly. It's kind of like the Uber app in a sense of like, you know, people are not going to be buying vertical SaaS in the way they were. They're going to be using a tool like warp to build their vertical SaaS. So you really want to be on the side of like, are you an app that enables other people to build apps rather than just like being an app? So I feel very good about our position there. And then I do think these cost dynamics will change over time.

20:35So right now you need to be you really need to be using like if you want the best result, it's like Opus 4.6 or Codex 5.3. And my hope is that Google enters with like a more competitive coding model. But I don't think that's going to be the case for very long. Like, I think you're going to be able to use it like an open weight model and get the same quality of performance. And for a lot of coding tasks, you just won't need to be at the frontier for much longer. And that's going to dramatically change the cost structure and hopefully get us back to a point where the best product wins rather than the cheapest product wins, which is kind of, you know, a little bit what's happening right now.

21:14now.

21:14Matt Paige:Yeah, it is interesting when anybody can build anything, the competitive, the positioning, the strategy element, distribution, all of that just becomes even more important. But the other point you mentioned too is interesting. The lines of what it means to be an engineer are blurring. You talked about product people being in there doing things. We have that in our company. They're in, you know, cursor and other tools doing things. Right. So that's kind of totally blending over too, which just, again, expands the, the base of potential users that use a tool like this. Um, and I'm going to, I like, I'll pull up just show you what I built with Warp.

21:58Matt Paige:And for those listening, we'll try to be descriptive. Um, but I was using, uh, Oz, I believe. So this is the new platform. And like you said, there's a lot of people using this. It was actually foreign to me when I first heard about it. So I first came into the environment. And for whatever reason, whenever I test out a new tool, I always default to make a to-do list app. I don't know why. It's the first thing I've ever heard. It's the hello world of agents. Exactly. It's a CRUD-based thing with a user interface. But I basically was conversating with it and told it to build it. And it was like talking through the stack and everything.

22:38Matt Paige:And I basically was like, do whatever you think's best. So built a plan. Uh, I worked with the agent and built it out and then I had to do this kind of cool, um, Apple, oh, liquid glass kind of vibe. Um, and like, uh, so, you know, build a to do list app on warp and like all this stuff, It just like natively did like the date and, you know, I can put these like different categories and things. You can add the task, but it's got this neat little, you know, kind of like, like I said, the liquid glass kind of vibe. But it's, you know, I had like maybe 20 minutes before the podcast and I was like, oh, crap, I got to test this out before we go try it.

23:22Matt Paige:And it's just mind blowing to me. And I'm a bit unique. Like I'm not an engineer by trade, but I understand how systems work. I used to work in data analytics with the theme of data scientists. So I understand how these things work, but I never crossed over into that world of, I am coding. But I mean, like I always wanted to, but then this transition that happened just like opened this whole new world for me. And I still think there's so many people that don't even realize this is a possible thing, which is great. We're still early adopters, I think, for this type of thing that you just did right here.

23:58I mean and it's just gonna grow and like I you know this is like a bold prediction that's out there but I actually think that like something like warp or cloud coworker like that's gonna be the interface that knowledge workers use for a lot of stuff rather than using all of these like individualized SaaS apps it's just gonna be like a personal agent control plane which is pretty cool that's where we're headed

24:28Matt Paige:it yeah i mean we've started doing some of that uh in in i think there's nuance between like oh i can go build a crm now well should you really go build a crm maybe not but like if there's nuanced cases within your organization that are unique to you that's not worth either going and buying something off the shelf or it doesn't really exist like it's just so easy to do that now yeah like for instance so we're automating a ton of stuff with oz and so we are i can just give you some examples i can even show you it's like we're automating all of our like social listening which was like we had a sass product for that but it was like the sass product wasn't very good didn't have any intelligence it was um and like you know the way that that should work is you just like run an agent that looks at what everyone's saying about you and like comes up with a report, keeps a memory of it.

25:21So you can see trends over time. At any given moment, I can ask the agent to do analysis on it. I don't have to worry about like, is this SaaS tool building like a graph UI for me? I just, if I want to do a graph on our social listening, I ask the agent to build me a graph and it builds me a graph. And so it's just this like, you can ask the agent to do things mentality it's a big it's a big mindset shift actually um but it's it's just a different way of doing knowledge work where rather than reaching for an app you reach for an agent who is more like a smart assistant uh you know open clause obviously in this space for like personal stuff but it's like you just are like okay i just want an agent that does this and and i actually think coding agents are like the um like a huge part of this because a lot of the capabilities that you want an agent to have are just like writing code on the fly um yeah so i think the whole world's

26:18Matt Paige:going to change totally if you want to show anything off feel free to to pull it up but um the one other thing too i always say this like context is the most important thing i feel like With sufficient context, you can do almost anything. But I saw this really interesting, it was actually, uh, some kind of video series or something. This guy was talking about this paper and it was actually a research study showing that your, you know, Claude.MD or your agent.MD like these markdowns. I'm following this. Are you okay? Good. It can be counterproductive. And it's like interesting. Actually, I pulled this up while we're talking through it for the audience.

27:02Matt Paige:I saw this the other day. I was like, oh, I should talk to Zach about this. But, you know, for context, there's Markdown files. It allows you to give the agents context for your project, how you like to approach things, really whatever you want to do. And it uses that for context, but it goes in a bit of hierarchy where there's like the ultimate system prompt for whatever tool you're using. then this kind of comes next and then whatever task you're doing. But what it shows is a, it overall, it can be about 20 % more expensive because it's taxing the, the tool in terms of tokens and whatnot. And then the other piece is the fact that it can actually steer it in the wrong way.

27:46Matt Paige:And like the perfect example, the person gave is like, if I say, don't think about purple elephants, what are you thinking about right now? purple elephants right yeah exactly so like the models kind of work in the same way you can tell it not to do something it can sometimes have a counter productive effect but i'm curious your take on this maybe skills context in general like what do you think yeah this i mean this result doesn't doesn't shock me uh to be honest it's like the way to think of context is it's it's kind of like a zero-sum type situation so you know these things have like a finite amount of context that you can put in them and if you're putting a bunch of context that's like kind of generic and a lot of it is being agent generated uh where it's like people will run like a slash init command and warp has this and we may get rid of it where it's like the agent will like add a bunch of generic stuff about your project like this is a typescript project and it's like that's just sitting there wasting space in the context window so that makes sense to me that it's like wasteful um and uh you know context is the sort of like the most precious resource that these agents have like when they go wrong and they go wrong all the time it's in my experience because it's they're either like forgetting meaning like stuff's been pushed out of the context window or summarized away or they just are like missing stuff that like a human would know and it makes them do things that appear to be dumb so it makes sense to me like if you're going to put something in the context window you should be pretty careful about it and these things like agents.md or cloud.md tends to just be like a lot of fluff because it's agent generated and it's not useful the better approach is definitely skills it's just basically context on demand like Yeah.