⚡️Warp 2.0: the Agentic Development Environment - Zach Lloyd and Ben Holmes

Watch the full show on YouTube: https://www.youtube.com/watch?v=0ZCMPodISoE

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:03Alessio Fanelli?

  Hey everyone. Welcome back to the Late in Space Podcast. This is Alessio, partner and CTO at Decibel, and I'm joined by Squix, founder of Small AI.

• 0:11Speaker 01

  Hello, hello. And we're very excited to, as Warp users, welcome two folks from Warp, Ben Holmes, who I've known for a while, and Zach, who's CEO and founder of Warp. Welcome to the pod.

• 0:22Speaker 02

  Thanks for having us. This is awesome to be here.

• 0:24Speaker 03

  Yeah, this is great.

• 0:25Speaker 01

  I don't even remember when I switched over to Warp. I think changing a terminal is a very personal choice for a lot of people. I think I was an I Term 2 guy for a long time, and then Warp came along. And you were one of the first to be basically the AI terminal. And obviously, you're like, woo, I want to try all the AI things. And like, it actually sucked. So that's how I introduced Warp. I don't know if you would augment it in any way.

• 0:48Speaker 02

  That's how we've been presenting ourselves. Like, we are, we are, today, I think, as this podcast comes out, releasing Warp 2, which, you know, we're positioning ourselves as being no longer just a terminal, but an agentic development environment. And I can kind of like, explain what that means. But short answer is like, we are, Warp already does and has done for quite some time, way more than what a terminal typically does, whether it's like, translate your English into commands, or doing tasks for you, or it has a collaboration platform. And so, you know, calling it a terminal, it boxes it in a little bit and creates sort of like, preconceptions around like, what you can do with the tool. The biggest new thing that we're launching though today is a

• 1:32Speaker 02

  coding agent in Warp, and it's like a state of the art coding agent. We should have some, I don't know the Sweet Bench score yet. Ben, Ben's been working on our, Ben's been working on Sweet Bench, but we're basically state of the art on Sweet Bench. I think we will, again, I don't want to be quoted here, or we can maybe edit this later, but like, we'll probably be number one, or close to it on T-Bench also, which is the terminal benchmark, which really we should be the best on. We're launching a, basically, it works very similar to Claude Code, which I know you all covered, and you know, I listened to that, that, that recording, where you can just tell Warp to build a feature. For you, tell Warp to fix a bug. And you can also tell Warp to do basically anything like you can tell it to help you debug why a server is crashing

• 2:17Speaker 02

  or why your Git is breaking or how to rebuild your Docker stuff. So it's it's a super flexible interface for doing these like agentic workflows. And so we're trying to break out of the terminal box to make sure folks get a good sense of its power.

• 2:31Speaker 01

  I want to say, even though like it wasn't like Warp one, it wasn't strictly like kind of an agent. So I just want to share this this experience I had, which I know that you saw. But, you know, I I've been I've been kind of critical of Warp, just as as everybody who uses the thing does. But then I was like, hang on. So like, I started installing this this new project called Home Assistant, which is kind of like an open source Alexa. And I didn't have some packages. And it just it just started suggesting things to me to fix my Python environment. All I did was I just hit Enter and it just hit Enter, Enter, Enter, Enter, Enter. And it mostly worked. There were sometimes I had to sort of go backwards. But this was agentic in a sense, except that I was powering it by hitting Enter. Like I was the human looper.

• 3:17Speaker 02

  This is my my favorite tweet you've ever done on us. I know we've we've blown up your memory several times and I apologize for that. But yeah, this this is like how most people I would say are who are Warp users are for the first time exposed to like, whoa, this thing can just sort of like solve my problems. That's like the most common inroad that we have for AI. So it was it was cool. It was cool to see you post that. And like, it's very magical when it works because you're like, it's the last thing you want to be dealing with as a developer. A lot of times is like, oh shit, like I'm missing this dependency, this path variable is wrong. And just like Warp can just fix that stuff for you. And so that that still definitely works. And I think we're the best at that. We've introduced the ability to let that go fully agentically now.

• 4:03Speaker 02

  So, you know, there's like a new set of, we call it like an agent profile, where you can, at the level of like, running commands or executing diffs or doing code search, you can control the autonomy level. And so I basically run Warp now with full autonomy. And, you know, as long as you're, you know, on like some production server or something, that that works pretty well.

• 4:27Speaker 01

  We should just let you go right into the demo if you're loading that up. And while you're loading that up, Alessio, I don't know if you have any questions that you wanna?

• 4:35Alessio Fanelli?

  I have one question. So since the ChatGPT desktop app, there's been the work with Warp in ChatGPT, which I used before. You can either do the SWX workflow of just let Warp suggest, or you can do ChatGPT work with Warp and ask it to figure something out. Obviously, you build support for that. What did you then decide to say, hey, we should just own the end-to-end agent and build the coding into it, and like instead of outsourcing the intelligence to a different layer?

• 5:01Speaker 02

  There's a few reasons. So the integration, I think you're talking about is you're using the ChatGPT desktop app. It can sort of like pull what's on your screen in Warp. The main reason is like, we feel like the integration can go way, way deeper if Warp itself is the thing driving the interaction, primarily because of like the type of tool calling that we can do like directly in Warp. Like when I think of ChatGPT desktop app or Claude desktop app, the big difference that I see with Warp is Warp is much more of like a read-write interface. And I tend to use those tools as like read-only in some sense. Now, you can kind of like MCP, hook them up and like do stuff through MCP with them,

• 5:43Speaker 02

  but the command line is really an interface that's built for doing. And so when you give agents the ability to run CLI tools or make code diffs directly within it, it's way more powerful. The other thing is like, we do a whole bunch of stuff in terms of like what the prompts should be for working with the terminal, how you get context correctly to do things in the terminal and like the, what we're now calling the ADE. And so it's not integrated enough if it's happening outside of our control and just using ChatGPT's prompts. And finally, we use a bunch of different models. So like ChatGPT isn't even the model of choice at this point. It's anthropic for us. Does that make sense?

• 6:23Alessio Fanelli?

  Yep.

• 6:24Speaker 02

  All right, let me start with this. If you are a Warp user, it's going to feel very, very familiar, except that the main thing that you see here is what we're calling this universal input, which is kind of just like a nicer input for driving either terminal commands or agents. One of the things that it does, so it will detect whether or not you are typing English or typing an agent command. So I could be like, tell me about this repo. And you'll see that this is like, I have this in a mode right now where you do auto detection. It realizes this is English. If I execute this, it's going to execute it as launching an agent. Let this run for a second. While it's doing this, the other thing is like, if you are a Warp user and you find this feature annoying, because it sometimes misfires,

• 7:09Speaker 02

  we now give you the control to just basically lock this into terminal mode or lock this into agent mode. I always use the auto detection because I find that that more useful. As a terminal, even if you're in auto detection mode, you can just like run regular terminal commands and it works the way that Warp has always worked. So you get all these nice features in the UX. And, you know, our thesis here is that the best interface for launching agents is actually pretty similar to what traditionally has been like a terminal UX. You know, you want this like grouped output. You want the ability to seamlessly run commands. And so we've kept a lot of that. The different things, which I'll go over here and show

• 7:50Speaker 02

  you, one are like the coding capabilities. So I'm going to go ahead and ask Warp here to fix a bug for me. This is a real bug in Warp. The issue here is like, if I turn this off, what should happen is that this model picker should go away. But right now, that's not happening. It's not listening to the setting. And so I'll just show you what the workflow looks like in Warp for fixing a bug like this. I will just be like, can you look up my linear issue? This issue is filed in linear already. 7252. Help me come up with a plan for fixing it.

• 8:32Speaker 02

  I'm gonna talk to Warp because I just find it easier and less annoying than like having like watching people have people watch me type. So I would say, please look at the attached screenshot to understand where the setting is. What I want you to do is figure out how to honor the setting. You should get a good idea from the bug description what the problem is. And so I just talked to it. And then the other thing that we've added is the ability within Warp to like give direct context to file references, which by the way, is actually very useful, even if you're just using Warp as a terminal. So I'll be like, you should look in

• 9:10Speaker 02

  at Universal in this file. And I'm gonna go over here. I'm gonna take a screenshot of this.

• 9:20Speaker 02

  And I'm gonna go over here. And if you've used other agentic coding tools, probably this is like like a fairly familiar type of workflow, but I want to show you some of the stuff that makes this cooler in Warp. So I'm gonna go ahead and do this. We'll wait here for a second while it does this, just to show a few cool things. So one is it's we now have full MCP support. So this is using the linear MCP. MCP in Warp is configured over in Warp Drive, or you can get to it through the settings. I have linear, I have notion, I have sentry. This is super useful. I assume people watching this are familiar with MCP, but if not, MCP is like a way of getting external

• 10:00Speaker 02

  context into an LLM-driven tool. A second thing that it's doing here is another thing that's new in Warp, which is we have codebase embeddings now. And I know there's been like some question of like, do you need codebase embeddings? It seems like there's differing opinions on this. We find that like definitely in Warp's internal repo, having the ability to do sort of like semantic search across your code files is actually quite useful. This is all controllable, I will say. So if you go over into Settings, you can see the codebases that have embedding, so that we can do sort of vector search over. You can control this and turn this off. We don't store any code. I know there's some coding tools that are actually like storing your codebase. We just

• 10:40Speaker 02

  store vectors for semantic search. If you go back over here into Warp, we'll see where this got to. So this found the issue in code.

• 10:53Speaker 02

  It gives me a nice view of what the problem is. It needs to check this setting. This is correct. And by the way, it's the way that I use these coding tools is I use, I don't really vibe code. I don't know how folks in this room feel about vibe coding, but I like to really try and understand, especially on a real code base like this, what it's doing. So this looks right, but I think actually it might be missing one step. So I'm going to tell it what to do. Make sure you listen for when the setting changes. You can follow the pattern that other settings use. You may need to call context notify when that happens. Please check that out and update the plan.

• 11:29Alessio Fanelli?

  What are you using with the speech to text? Just the Mac OS?

• 11:33Speaker 02

  We're using Whisperflow, which is another startup, which has got like a really good API around doing speech for text. And I do find the speech for text, when doing this kind of prompt driven development is like a magical thing. Otherwise I'm just like typing these long things. While this is going, I'm going to go over here and fix another issue. So this is one of the cool things about using something like the terminal form factor is that you can naturally multi-thread while doing this work. And actually I can see notifications going on. If I come back over here as this agent is working, I'm going to start over here. For this, we have a null pointer exception, annoyingly, in our go code base that is being

• 12:16Speaker 02

  tracked in Sentry with this issue. So I've kind of like, you know, set this up to do this. I'm going to ask Warp at the same time to go ahead and try and fix this for me. So I'd be like, could you look up this Sentry issue? Doesn't get that perfect Sentry issue. And diagnose why there's a null pointer exception. I'd like you to look through the code, figure out where it's happening and suggest a plan for fixing it. Once we have a plan, I would like you to create a new branch and on that branch, let's go ahead and implement the fix, assuming the plan looks good. And then we will write some unit tests for it as well.

• 12:51Speaker 01

  My god, the amount of typing and work that you put in here is just an order of magnitude above anything that we normally do.

• 12:59Speaker 02

  So this is gonna go do its own thing. Over here, you can see this one is done. I want to show one other cool thing that we've added, which over here is the ability to see the status of multiple agents. And so agents in Warp are just sessions, they're terminal sessions. They're running within terminal sessions. You can do them in split panes. You can do them in tabs, anywhere you could run a command, which is like analogous to running an agent, which is an intelligent task. You can run one. And so this gives you an overview of all the ones that are running in addition to what their status is. And so what we're trying to build towards here is a world where you're truly doing this like multi-threaded development, where this is going and you have multiple of these things going at a time

• 13:43Speaker 02

  because you don't really need to sit there and watch them. So let's see this. Okay, so this worked. This took my advice, which is nice to see. Warp's built in Rust, by the way, which is why you're seeing Rust code here. One thing that's very cool, Warp is over a million lines of Rust code. And this is like, up until very recently, this type of agentic stuff just did not work in Warp because it's not built on any public UI framework, like we have our own internal UI framework. And so it really needs to understand our codebase to do this. So I'm gonna be like, I'm gonna ask it to go ahead and make the change. And then we will go do something else. So I'll be like, yes, please make, I'm gonna talk again,

• 14:26Speaker 02

  this change for me, but please first create a new branch with the appropriate linear branch name format so that it's tied to the linear issue. So this is how I code now. I'm curious if other folks go like this. It's pretty wild. It doesn't work perfectly. Like, you need to be able to go in and edit. And I'm gonna show you that next, how that flow works. But in the meantime, for demo, I'm going to show you I can use Warp to do production work. That's another kind of agentic workflow. So here I'm going to use a feature of Warp that's called Save Prompts. So we have a save prompt that I made before, and a save prompt is just like, it's like a paragraph telling

• 15:09Speaker 02

  an agent what I wanted to do. But the thing is, it's reusable. And this is also shareable across your team. So Warp has a bunch of features that are really cool for supporting teamwork in this area over here called Warp Drive. This is something that if I want to do logs analysis, I can pull up. I'm going to go ahead and start this. And what it's going to do is it's going to help me get some logs from our Cloud Run instances, but it's going to walk me through that. And again, you can see over here, I think this is super cool. You can see what these other agents are doing. Let me go back over here for a second and see if it started adding code. So it has. So while this one is running, I want to show you what the code experience looks like. So you get sort of similar to Claude

• 15:54Speaker 02

  Code here, which I think is the paradigm that we're most similar to. Okay, this one is waiting for my attention over here. You get these inline diffs. One thing that I should show you is that I also have the ability to edit these. And so let me go over to my AI settings for a second and turn off some of this autonomy. So I'm going to change this to the agent should ask me for a code diff. And I should, oh, it already made the code diffs. Hold on, I'm going to stop this for a second. So I want to show you what this looks like. Could you add some comments into the code diffs that you just created, please? Okay, so I can do this. Let's see how this one's doing. So this one now needs

• 16:36Speaker 02

  my help. Oh, I got it over here. Okay, amazing. So What's cool about Warp compared to say Claude Code or Aider or Codex CLI is that as you do this sort of like linear agent-based coding workflow, you can actually use like some elements of rich UI to do it. So this is a diff viewer, right, that we have built right into Warp. If I want, I could also go into here, and this is a full code editor. So this is brand new with the IDE launch, and I can sort of, you know, actually edit in here if I want. My feeling is that you're going to be doing fewer and fewer manual code edits. And I know, like, again, I watch you have the Claude Code folks on, I think

• 17:20Speaker 02

  that they agree. But sometimes you need to do them. And not having to context switch to an editor outside of the tool, I think is a really nice piece of the workflow. So I'm going to cancel out of this. This one also, you can see, is now paused. It's the same thing. So I asked it to add comments, start to ask comments. I'll go ahead and apply this comment, but I'm going to turn Autonomy back on so that we can, it doesn't have to ask me all the time. Okay, over here. This one, in the meantime, is waiting for me. I'm going to go ahead and try to let this one run to completion, just because I don't know how long it should take with this demo. But I'm going to, remember, this one is doing some logs gathering. I'm going to be like, can you use warp-server-staging,

• 18:04Speaker 02

  use the warp-server-cloud-run instance, and then pull all info, warning, and error logs for the last 10 minutes, do an analysis of the most common categories of errors for those three log levels, and print them out for me in a nice format. So if you've done any like, production work in your career, this is like the last thing that I want to spend my time looking up, like figuring out how to use these tools. Like, it's like kind of hell. And this is like a place where I think Warp really, really shines. So it's doing these commands. And again, a really nice thing here is like, because Warp is at its core, sort of terminal UX, you can always drill in and like pop these open.

• 18:50Speaker 02

  See the result of the tool call. It is, let's see if it can figure this out. So yeah, it's figuring out what all my services are. It's also, this is cool. So it's using this little derive from thing here. Means it's using one of our rules. And so this is very similar to like cursor rules. The biggest difference that these are, is that these are sort of like cloud synced. And so you can sort of use these wherever you log in, they're sort of shareable across team. And so like, this is our version of of curse's rules. I think we will also, by the way, probably just literally support the cursor rule format. This one already kind of completed. So this is a peek inside our staging, like where do we have like,

• 19:32Speaker 02

  what are we doing? These are our GraphQL operations. We see Anthropic. We have warning level logs here. We have some timeouts of Gemini because we're, we're using a whole bunch of different AI services under the hood. Some 500s. I just think this is like so awesome. Let's go back over to this one. So I could be like, you know, continue fixing the NPE in the go code. I'm gonna, I think I basically hit most of what I want to show. I'll show one last thing, which I think is a super cool and very differentiated thing, just, just to warp. One other place where AI is really helpful, which is not something that's super well known in warp, but it can be really powerful, which is like, if you go into a, um, a REPL.

• 20:18Speaker 02

  And here I'm inside a Postgres REPL. You have AI access even within this REPL to do natural language into REPL command translation. For instance, I could be like, show me all tables and it will live translate. That's very trivial, but I could be like, you know, I, uh, oops, I hit the wrong key. I could be like, show me our users table. And I don't want to do anything, I don't want to think too sensitive here. Maybe I'll do like, show me our workflows table.

• 20:50Speaker 01

  Yeah, careful.

• 20:50Speaker 02

  Try that again.

• 20:51Speaker 01

  Demos on your production data.

• 20:53Speaker 02

  Yeah, I mean, this is not real data. So it's like, I think fairly safe, but even showing the like schema like maybe scared me slightly. But the idea here is like, you can then be like, you know, show me all workflows related to Git or something like that. And what's cool here is like, it is generating SQL for me on the fly and also learning the context based on the contents of the session. So if you're doing like work in Node or Python, Postgres, MySQL, it works in any REPL. And it's a, it's a pretty differentiated and cool feature of Warp. So I think this is a good demo.

• 21:33Speaker 02

  You know, what I was really hoping to show is like Warp can code and Warp code is at a level that's like comparable to Claude code. I think it's actually better than like Cursor's agent. It's probably better than Windsurf since they aren't, they don't have Sonar for. It's a really, really great coding agent and a really awesome and differentiated UX for doing coding with the inline editing features that we have. It's also an awesome interface for agent multi-threading. So you can have multiple of these running at a time. You get these cool notifications around what the agents are doing. It's got this universal input around like how you control the agents, including with voice, which is super useful. So hopefully this gives everyone a good sense of, you know, what we think the future

• 22:16Speaker 02

  of development, like an agentic development environment should look like. So let me come back to the session and stop sharing. How was it?

• 22:26Alessio Fanelli?

  That was great. As an early Warp user, as I was telling Ben, I'm still using pound sign, ask for it, I commend.

• 22:32Speaker 02

  I know, there's like way better way of doing it.

• 22:34Alessio Fanelli?

  How are you thinking overall about like the UX and the shape of it? You know, I feel like even now you have, the Warp drive was kind of like there on the left, but maybe it's almost more interesting to have the different, see the different agent run times run versus hiding them. Like what are like, maybe someone like the baggage that you're bringing with you, you know, and like where you see things going?

• 22:56Speaker 02

  The different like primary thing that I think we want to feature in the UI is like, what are your agents and what are they doing to your point? And like, right now, you can sort of see that in that tab thing up top, but it's also kind of like what a tab is or what a pane or what a session is. And so we are, we're trying to think through on the design side, like, you know, what, what would a agent first interface look like? We're also trying though to um, kind of like build a bridge for people who are already using Warp. And like, there is like some baggage there, like, here's a hypothetical way we could have built the app. We could have made it look like Slack or something where you're like,

• 23:39Speaker 02

  you know, your primary modality is like you chat with an agent. I think that that would have been like maybe a step too far for current users. It's debated and I bet I'm curious if you have a take on this, but like, I think there's something awesome around like, hey, you get like the familiarity of the terminal form factor. You get this very, very clean abstraction of like tabs and panes and sessions are containers for agents. But maybe like the tab bar or like the agent switcher like that. There's some way of combining those that we haven't quite settled on yet. Uh, Ben, what do you think?

• 24:14Speaker 03

  Because I did get to work on some of the status features in the tab bar and also the management pane that you see in the corner. And around that time, Claude Code was experimenting with like, obfuscating the details of what it's doing beyond a nice to do list that shows it checking off interesting features, which is similar to our management pop up as well. And I've found myself enjoying Warp's approach a bit more because as I'm sort of feeling out the style of a PR or if I need to make edits to a code change, I never have to open an editor to actually review the files and get diff everything. Cause Warp did the first step of we'll show the agent output. We'll show the diffs that are occurring. We'll have an option to edit it.

• 24:57Speaker 03

  And one thing you didn't demo was also a little fast forward button you get at the bottom where, yeah, if you turn off automatic mode for everything, you You can hit a fast forward button whenever you want it to just go off and do everything. Otherwise, it will just notify you with the desktop ping whenever you need to review a file. And that just feels like the natural next step from using like tab, tab, tab inside of Cursor with Super Maven. Where I'm still looking at the code, I'm still working with it. And I have an option to fast forward the more I trust this tool. And then eventually I may just check the management pane and my desktop notifications and rarely have Warp open at all. Except for inputting prompts.

• 25:37Speaker 03

  So it feels like a more progressive interface since we own everything beyond like the CLI itself. It's all the tab views and everything.

• 25:46Speaker 02

  Yeah, it could be that that that real estate on the left is like prime real estate where agents should go. It's interesting. It's an interesting question. We do have the historical like warp drivers over there.

• 25:56Alessio Fanelli?

  It's almost like where's the strength of it, right? I think like if you really take it further, it's like, well, if Warp code is so great, I should just pay for Warp code to run in Ghosty as well. You know what I mean? It's like, it's interesting. It's like more of the AI piece comes into it and like the intelligence itself. It's like what's valuable to me more than like the terminal itself. Why people like Claude Code is they bring it around, you know? And like with Warp Code, it's like, I would love to bring it around, but you can imagine somebody saying, I'm not getting anything from Warp outside of Warp Code. So why do I have to only use it in Warp? You know what I mean?

• 26:31Speaker 02

  Doing like a headless like CLI version is interesting. My take on that is that there is a lot of value in what we can do with the UX. That if it's a pure CLI app, you just can't do. So like, there's no way for Claude Code it running in Ghosty, let's say, to have an editable diff. Uh, now you might say like, well, why do I need an editable diff? Like if it does it perfectly, even if you don't want to edit the diffs, I think honestly, like the next thing that we'll probably be building is like, basically a way in Warp of code reviewing what the agent has produced. Cause I think that you totally want, like if you're having an agent write code, I find it very, very cumbersome to like push all the changes up to GitHub, review it myself, then come back.

• 27:16Speaker 02

  Back into my tool and have the, you know, then retell it to do it. And so there's all these like native UX things that I believe you want. If your primary workflow is like launching and managing agents, there is not available if you're a CLI app. I would say like, our likely path forward on that is like, we will probably at some point make a Warp CLI app. Where if you really want to run it headless, like you want to run it in CI, because I do think this is an advantage of like Claude Code. You can run it there. But I think for like the hands-on keys, like I'm a developer coming into work and I want to like, use these tools to delegate work. I think it's a much nicer user experience. If you're doing it in Warp than if it's just running a CLI app.

• 28:00Speaker 02

  Which is, by the way, it's like, it's kind of the thesis behind Warp itself, which is that like, just using CLI apps was really hard, a pain in the ass. And like, having like the mouse work and that kind of stuff is, is nice. And so, you know, I, I think it's in keeping with that original philosophy of like, how do you build a super accessible user experience?

• 28:20Speaker 01

  I don't know if, uh, how big of a question, you can probably decompose this question. But as a comms person, I think I have trouble fitting Warp in the universe of things. And because you're a bundle right now, you're a bundle of UI, terminal, and agent. And, you know, what if I really like Claude Code? I'm gonna try to use Claude Code inside of Warp and actually I've done that. But you've basically made a competing harness that I now have to think about, you know, using. And then obviously, you know, before Warp had Warp 2.0, I actually did the voice coding that you just did. Inside of Windsurf, I use Super Whisperer instead of Whisperflow, but same thing. I just talk into it and did that.

• 29:04Speaker 01

  And then I think the last piece for me that Warp doesn't have yet, but I presume it's, you know, if you, you might eventually explore this, is some kind of sandboxing, some kind of cloud background sandboxing. Because you can only use Git for so much, like, it's gonna have, the code has side effects, terminal commands have side effects by definition. And they're gonna interfere with each other. So like, basically, like, I think like, where are you going to own and where are you gonna compete?

• 29:28Speaker 02

  It's an awesome question. You might have to like, I'm gonna try to answer all these. You might have to remind me of all the different parts of it.

• 29:34Speaker 01

  Yeah.

• 29:34Speaker 02

  It's like a complicated question.

• 29:36Speaker 01

  Agent Harness, UI, and Sandbox.

• 29:39Speaker 02

  I think the Sandbox one's the easiest, which is like, we're local right now, like in a very similar way to like Cursor Win Surf. And I think that's the right place to be right now, but I don't think it's the right place to be forever. The reason I like it right now is because it's where developers are doing their work. And when agents fail or need to be corrected, it's very convenient to correct them without the like, long feedback loop of like, either SSHing or trying to like replicate a remote environment locally. So this is a much faster feedback loop in a period where these agents are pretty fallible, right? So they're not perfect. So that's why I like it local now.

• 30:19Speaker 02

  But if our thesis is that multi-threading is going to become really important, it's gonna not be doable on a single machine. Like if you, if you didn't like our memory usage before, you're not gonna like it when we're running, you know, 10, uh, 10 agents. Uh, it's just not gonna scale.

• 30:35Speaker 02

  So there's like a resource part of it, and then there's like a sandboxing part of it, which I also agree with. I also think like, the sandboxing thing is very cool, but a very big barrier to adoption. In the developer tool space, we've been trying, I think, forever to get people off of their local machines to do development, whether it's like CodeSpaces or GitPod or whatever. It's such a frickin' pain to set up. It doesn't work for all kinds of development. If you're doing anything with local hardware or mobile development or whatever, it's like, it just is not ergonomic. And so, you know, we'll get there eventually. I've, I've been saying that forever, but I, I think right now you gotta start where the development happens. So that's my answer to that one. What was this? The, the first question was like, what are we? Like where in the landscape are we? Something like that.

• 31:21Speaker 01

  The first part of the stack was Agent Harness, but basically, yeah, that's the real question, right? What will you own versus what will you like sort of collaborate with?

• 31:29Speaker 02

  The way I think about this is like, we've been a terminal, right? So like the developer tools landscape, I think for a long time has been like, are you a terminal? Are you an IDE? Or maybe are you like a CLI app? Or like an IDE extension? I guess would be the other like thing. What our thesis is, is that the primary workflow of development is changing. And you know, it's changing from your primary development workflow being like, hey, if I want to make a coding change, I'm going to open an IDE. I'm going to open a bunch of files. I'm going to hand edit them. Then I'm going to open the terminal and I'm going to like run a bunch of commands. And I'm going to do a bunch of stuff with Git. I'm going to do some stuff with Docker. I don't think that's how developers are going to be working.

• 32:14Speaker 02

  Like, I think the way that developers are going to be working is primarily by prompt. Instead of like, if I want to make a change, I'm not going to open an IDE and open five files and use autocomplete and LSP and all that stuff. I'm going to ask an agent to make a change. If I want to do something in production, I'm going to ask an agent to figure out like why my servers are failing. And you know, this may sound like out there, but I think that you want a new type of workbench for that workflow. And that is what Warp is like aspiring to be. We don't fit cleanly in any of those boxes. I think that that like hurts us to some extent because it's like, if I'm a lifelong developer and I'm like, okay, I get it, CLI app. Or like, okay, I get it, IDE.

• 32:58Speaker 02

  We are trying to be a different kind of thing. And being just like a terminal, I think is like, doesn't do the tool justice. It's not like, it's just like not what it is anymore. And so we're trying to build that next thing, which is the most natural way of supporting this like prompt driven development workflow. And we're coming at it through the terminal form factor. That's our DNA. It's like, that's the app that we have. I actually think it's like pretty close to the right form factor for doing this kind of development. I think it's certainly much closer than the IDE. Like the IDE is like a static view of all your files at a point in time. And like, if you look at all the IDE agents, they're like in a chat panel, they're basically Rebuilding something that's kind of like a terminal.

• 33:41Speaker 02

  I think Claude code is like proving that this like prompt driven workflow is super valuable and feasible. But I think from a, I just think that you're going to run into a wall if like your entire app is a a terminal app, a CLI app. And I think even hearing the Claude code folks on your podcast, like I would not be surprised if Anthropic like launched a thing that looks a little bit more like Warp, where it's like a harness for running a whole bunch of different Claude codes, but it's an actual like thing that can live in your desktop. You know what I mean? Cause I think it's, it's like from a distribution standpoint, it's pretty limiting to just be like a terminal app. Although there are advantages to it.

• 34:18Speaker 01

  I think it's a very difficult question. I think you answered it as well as you could. I think it, it's, it could be cleaner, but you know, sometimes you just have to compete where you compete. I think that the challenge for you is just the sheer amount of firepower. The big labs are putting behind their agents, right? And you're coming out saying you, you got a better agent in them. Good for you, but also like, you know, you're fighting them.

• 34:43Speaker 02

  So it's an interesting position because we are competing with them and we're also their customer. And it's, it's an interesting question to me from their perspective.

• 34:57Speaker 02

  How much do they want to own the end user relationship? How much do they want people building on their APIs? There's this interesting dynamic where I think Anthropic is like pretty clearly the leader in the, in the coding models right now. But I don't know if that's going to be perpetuity. Like they have a competitive dynamic with each other also, right? Whereas we have this interesting advantage of like, if Gemini became the best coding model in two weeks, everyone who had like gone all in on Claude code would be like in a little bit of a weird position. Whereas we're just like, okay, we'll make Gemini our default model. So there is like some advantage to being where we are at a layer above.

• 35:42Speaker 02

  You know, I'm sure that the model companies are thinking like, how do we lock people in? Cause they don't like, that's a very big risk for, for Anthropic in my opinion. To just be like, okay, like Gemini is best next week. Cursor moves, like cursor, I think is some very significant portion of their revenue. I don't totally know, like, is the short answer, but the sort of north star for us is like, can we build the best user experience around these, like, around the models? I do think there's a bunch of alpha in terms of like, what's the UX like? What's the prompt engineering like? What's the context retrieval look like? But yeah, we're competing with them and it's, they have more money than us, that's for sure.

• 36:23Speaker 02

  They have more people, they have a bigger brand, like, they have a bunch of advantages that we don't have, but we have some advantages too. We actually have a lot of developers using Warp, by the way, which is a very significant advantage. It's a good thing for us that we started building up that user base a while ago.

• 36:38Speaker 01

  Do you have any public numbers that we can reference of like, what are you saying regarding growth trajectory of Warp? Just to establish in people's minds if they've never heard of you.

• 36:49Speaker 02

  You know, it's like almost 600,000 active users on Warp right now. Our revenue since the beginning of the year is growing like, depends on the week, between five to 15% every week. And so like, it's actually a pretty fast revenue ramp. Now, we're not nearly at the scale, just to be clear of like Cursor or even Windsurf, but like, we have pretty awesome product market fit with people who discover the coding and paid parts of Warp. So the other is, we're in an amazing market right now, is the other thing I would say is like the, and this wasn't true a year ago, but like, now it's, it's very expected to pay for developer tools.

• 37:32Speaker 02

  And for the first time, we have companies that are coming to us being like, my developers aren't using the latest AI tools, like, how do we get them to use it? And that's like a totally new development, which is very good for us. So we're growing really, really well and we have this very big existing user base. You know, not big like relative to like ChatGPT is like 500 million users or whatever, but for a developer product, like pretty substantial and that's also growing pretty quickly.

• 37:58Speaker 03

  Yeah, and usage of it has also been going to the right from dipping your toes into just asking Git questions to actually asking it to do long running tasks. That's the thing we've been looking at, but all this growth has just been on the back of, I want something to help me through Git and Vim, which is crazy. And now it's able to do a lot more things, asking through the exact same input box. So people are just finding it organically and now they'll find it a lot more readily.

• 38:23Speaker 01

  While we're on the topic, what is pricing changing at all for this new thing? Have you looked into any other models?

• 38:29Speaker 02

  We're giving away more, basically.

• 38:32Speaker 01

  Okay.

• 38:32Speaker 02

  The two plans were a, actually, there's three big changes. So there's, the two plans were the pro plan and the turbo plan. Pro was a thousand, that's going to 2,500 requests. The turbo plan is going from 3,000 to 10,000 requests. And then we're introducing usage-based pricing. So if you hit the limit on the turbo, you can buy more credits. We're trying to also get that into the other two plans. I will say this on the pricing is that it's like, it's very hard for us to figure out how to price this correctly. Like, I feel like a lot of the, a lot of the AI, like, what we're trying to do is like, we really want people to get a ton of value out of it, but it's very expensive also for us, right? And so we don't know if it should be a subscription, if it should be usage-based.

• 39:16Speaker 02

  I see pros and cons to both approaches. It looks like our competitors also don't quite have this figured out. You know, if anyone listening is like paid warp customers, it's just like, we're, we're, we're, we're trying to, to, to do this right. And we're kind of erring on the side right now of like, let's let people use it as much as we possibly can. But it's tricky thing to sort of price right.

• 39:37Speaker 03

  And we also have the benefit of having access to every model. So I know you can fall back to Gemini Flash and use that for simple ass and you don't have to worry about overages in those cases.

• 39:48Speaker 02

  Yeah, we give users control on that.

• 39:50Alessio Fanelli?

  I think that's one of the things that people don't want to think about. I think even in Cursor, you have like the model picker and people are like, hey, look, I just always want the best model because I'm paying the monthly thing. And that's why even with Claude Code, when it first came out, I think it had a way less adoption than it should have because it was only usage-based pricing. So people would do something and it was like, you pay five cents and they would be like, okay, I'm over it. And, and now that you can use the cloud subscription, people are like, oh, whatever, I'll spend a hundred bucks a month on Claude Max and do this.

• 40:22Speaker 02

  I don't know if you want to go into the business model. The challenge with the subscription is that, um, you know, for like a $50 a month subscription for warp turbo, we probably have some users who are costing us like a thousand bucks or something. And we have some users who aren't fully utilizing it. And so to price that correctly is very. Is very challenging, but you're absolutely right. Like the, the reason we haven't done the usage based, just the pure usage base is it feels psychologically like it's like in your face of like, it encourages like rationing the usage in a way that I, I think is, doesn't let people get the most out of it. So I, I don't really know how to, to price it, right? It's been a real fun, uh, business challenge.

• 41:05Speaker 01

  Well, uh, I think the orb guys, you know, have, uh, they published an online track talk at, uh, the New York A I E where they have some thoughts on pricing. But yeah, it's something that I think people are actively trying to figure out. I think this is honestly a fun, like economics, uh, challenge for any nerds out there. They wanna, wanna get into the pricing and talk to Zach. Just one more thing on the team side or enterprise side, you know, like, uh, that was basically the original pitch for Warp. Any enterprise AI needs you're, you're seeing that maybe individual developers wouldn't, wouldn't have visibility into until, until they hit that size?

• 41:39Speaker 02

  Just to unpack that question a bit. So the original sort of business model Warp has existed before, uh, L L M's became really good. The original business model was building a collaboration platform around the command line where you would share workflows, you would share notebooks, you could share sessions, all that type of stuff.

• 41:56Speaker 01

  Postman, postman for CLI.

• 41:58Speaker 02

  That's the easy way to name it. Post, postman for the terminal. And, you know, we do have, uh, paying customers who, who use that very extensively. It's actually really, um, it's complimentary to the AI in the sense of like, all the knowledge that goes into that becomes, goes into an embedding space that the, uh, the AI can access. And so if you have like your like onboarding guide in there, or your like firefighting guide, or your database migration guide, what's really cool is like you can just ask AI and it'll get it. Now, it's not a complete solution. And like the complete solution is like that plus MCP. And the other place where I think we can add real value is by making MCP configurations

• 42:40Speaker 02

  shareable, by making rules shareable, like that type of thing. And having the team concept also makes it easier for us to do enterprise wide, sort of like admin controls, which is honestly the, probably the biggest thing that they want, is they want like, they want to enforce that everyone's on a team, that they sign in with S S O, that they use a certain set of models, that there's like different data retention policies in place and that sort of stuff. And so the team construct. Lets us do that. And like, you know, we've, we've gone through all the work of like, being SOC two compliant and like all the stuff that you need to do to sell the enterprises. And we have a bunch of enterprise customers, which is a big part of our future business.

• 43:23Speaker 01

  You have all the workflows, you have all the, the keys and everything, right? So you actually own a very valuable part of that information. I guess my question is, is there a standard gateway that people are using where they, they wanna, they're like, okay, run every AI request through this thing, we'll track everything. Because I, I need to be on top of my team's AI usage.

• 43:42Speaker 02

  That's another enterprise feature where it's like, they could be like, I don't know exactly, I, I'm not sure if I'm understanding gateway right, but they might be like, we have a internal Azure based OpenAI LLM endpoint. And we don't want you going to the public endpoint. And so we can support that kind of configuration. We can't support like an on-prem type configuration, which some people want.

• 44:06Speaker 01

  Everywhere enterprise wants on-prem.

• 44:08Speaker 02

  That's been like an impediment at some places, but not, it's not a huge one at the moment.

• 44:13Speaker 01

  I'm exploring for like, what is the state of industry right now? It seems like it's not as secure as it could be. And I see some potential for other founders to like get in there and like do stuff.

• 44:25Speaker 02

  Ah, interesting.

• 44:26Speaker 01

  It's very obvious where all this will go. It's just no one's built it yet. So.

• 44:31Speaker 01

  Yeah, I, I, I think we're, we're, we're ready to start wrapping up. We had some miscellaneous questions if you're, if you guys are up for it.

• 44:37Speaker 02

  Yeah, let's do it.

• 44:38Speaker 01

  The first thing we saw, both Alessio and I were texting each other. You're a production GraphQL user. I haven't seen one for four years. What's your take on GraphQL?

• 44:51Speaker 02

  Oh, yeah, I, I don't think, I'm gonna let Ben take this one. I, I'm not coding much against the GraphQL. What do you, by the way, what are you seeing out of curiosity? Just so I know how to go back to our engineering team.

• 45:03Speaker 01

  What do you mean? Oh, TRPC, like old school APIs, I don't know, like.

• 45:07Speaker 02

  Just not GraphQL. Just not GraphQL. Ben?

• 45:09Speaker 03

  Oh, man, that's tough. Cause I know that more of the complaints about our backend are the fact that we leaned on Golang than the fact that we leaned on GraphQL, which is surprising. Mostly because the way we model everything with agentic tool calls doesn't match to Go struts very well. So that could be a cautionary tale. I don't know. But GraphQL, yeah, because I come from literally building a framework that included RPC endpoints. Like I featured that a while ago. And it only works if you have the full stack written in the same language and same contract. So I don't know, I feel like for tool calls, it kind of makes sense because you can define the structure and return different data. But if we switch to REST tomorrow, I feel like it would still work.

• 45:54Speaker 03

  So I don't know if there's a spicy Twitter take I need to get out of our engineering team, but I think it's fine.

• 46:00Speaker 01

  It's not about, does it work? But like, does it bring the benefits that you expected from the extra level of extraction? Right? There's a cost, fine, but like, there should be a benefit.

• 46:11Speaker 03

  I don't know, because I know at the last company that I worked out, the benefit was not there. It was over engineered. And here it feels the right amount of engineered.

• 46:19Speaker 02

  You could be honest if you've been like.

• 46:21Speaker 03

  I've been working on the front end since joining. That's the tough part.

• 46:25Speaker 01

  Yeah, he's the consumer. Consumers are great.

• 46:27Speaker 03

  Oh, yeah, as a consumer.

• 46:28Speaker 01

  It's a producer of the GraphQL API that sucks. Like, anyway, okay, I'm happy to move on.

• 46:35Alessio Fanelli?

  I think like, there's also the more history of the company, kind of like getting started, the first pitch, pitching Dylan Field. I know Dylan has been very active also in the AI space and kind of talking about that. Any fun backstories that you haven't shared?

• 46:50Speaker 02

  I mean, so the original, yeah, the original like pitch and vision was, hey, there's this like tool that most developers use every day, which, you know, the terminal that I think is like kind of hard to use tool, hasn't really changed in a long time. But if you get really good at using it, you can be like an incredibly productive developer. And like all the best developers, or a lot of the best developers I've worked with that are really good at using it. And so it was like, can we build a better version of that as a product? I think from like a business standpoint, it was like, you know, there was like a lot of question marks. I think I got like, good amount of trust from our, our investors based on, I guess, kind of based on my background.

• 47:33Speaker 02

  So I was a, I used to be the lead engineer on Google Docs. I was the principal engineer at Google and. I think that they gave me a bunch of trust. I think it was also like a weird moment in history where it was like COVID and pre LLM. From an investor perspective, I think it was like a weird like, what do I invest in at this time? Do I do more SaaS? Like, people were trying to like, pattern match. They're like, oh, there's no collaboration platform for developers around this tool. And like, Figma was like, you know, successful with adding, you know, making design collaborative. And so I think there was a bunch of that going on. But, uh, yeah, I don't know. I don't have any like, super specific fun story, other than that, like Dylan is awesome to work

• 48:17Speaker 02

  with. Uh, and has been like, super big supporter, especially as things have changed. It's interesting also, like, I think Figma honestly is now like in the developer space. Like, I don't know if you all have tried any of their tools. Like, they're further up the, up the like, app building kind of stack, like competing with more with like lovable and Bolt. But it's, it's like every company is like building for developers now. It's pretty wild.

• 48:40Speaker 03

  We have MCP support. Striding it out. It's really cool.

• 48:44Speaker 02

  Yeah.

• 48:45Speaker 01

  My Dylan anecdote is I ran into him. I was like, what did you see in, in Zach that like, you led like a $50 million round, uh, you know, in a terminal? And he was like, you know, this guy's extremely cracked. Like, he wasn't specific at all. It was all by his race.

• 48:59Speaker 02

  Yeah, I guess, I mean, I guess Dylan thinks highly of me, which is a big compliment because he's, he's a, he's a super, super smart guy. Uh, so that's nice.

• 49:07Speaker 01

  He's basically never wrong. He's just early. And like, that's fine as an investor, like.

• 49:13Speaker 02

  But having like an investor who's a, um, who's an operator who runs like a real company, like such an awesome like cheat code compared to our institutional investors are amazing too. But like, having, having Dylan to ask questions to is like incredible.

• 49:29Speaker 01

  Yeah, but he's like a, like, I've never seen like an individual person lead like that until until he did. Too much to chat on maybe back on stuff. Any parting thoughts, call section for people as we wrap up the episode?

• 49:43Speaker 02

  I mean, just the obvious one. I'd love people to try Warp. Try it with like a new set of eyes. Try and do stuff in it that you might not have tried to do before. If you're an existing user, if you are a new user, please give it a shot. Give us feedback. It's at warp.dev. It's free to try on every platform, Mac, Linux, Windows. It even runs via Web Assembly in the web. If you want to like get fancy, try it, give us feedback. I'm really, really psyched about Warp 2. I want to hear what people think.

• 50:11Speaker 01

  Ben, I know you're a few months into Warp. You've done a bunch of stuff. You had a great talk. What's your plug?

• 50:18Speaker 03

  I mean, one thing I used it for was just to learn some new languages. I was trying to use it to build a Swift app because I touch Xcode with a 10-foot pole. I just do not want to deal with that. And it's totally good at writing like macOS desktop apps in Swift. I threw it out on Twitter and my hot take was, I don't know if we need things like React Native in the future because you can just ask these LLMs to translate between different native environments. And people said the training set's not there, it won't perform. And sure enough, Warp seems fine at holding Swift or Rust or Go or anything you're trying. I think that's the big thing is it's really good at languages outside of just JavaScript and React, which seems to be the zeitgeist for all these other tools.

• 51:02Speaker 03

  But I've been using it for everything but JavaScript.

• 51:05Speaker 02

  That's a great point. That's actually not our special sauce at all.

• 51:08Speaker 03

  Although it's very good at it. Oh, you found it. Yeah. It riled up the React Native crowd for sure because, well, React Native is a really great tool.

• 51:19Speaker 02

  It's got a lot of engagement, Ben.

• 51:20Speaker 03

  Yeah, I know. Sometimes you have to throw out the weekend takes. Otherwise, yeah, there's my Twitter, beomzdev. I do whiteboard explainers on all sorts of technical topics and I put Warp in there wherever I can. So, check them out.

• 51:32Speaker 01

  We're going to try to amp up video production for Lean Space and I'm going to try some whiteboarding. Not to your level, but like, I think we should draw things. You know, I think there's like enough podcasts where it's like four guys on a Zoom chat.

• 51:45Speaker 03

  Or four guys on an Excalibur. You need the physical whiteboard. It's just different. It feels like you're inviting someone in.

• 51:51Speaker 01

  Okay, well, thank you so much for joining. Thank you for working on Warp. We're both, you know, happy users. I'm sometimes grumpy about it, but I only complain about things I use, right? So, you know, um, thank you for, thank you for working on Warp.

• 52:02Alessio Fanelli?

  Thank you, guys.

• 52:03Speaker 02

  Thank you, uh, thank you so much, uh, for having us. This was awesome. Really appreciate it.

• 52:08Speaker 03

  Yeah, it was great.