What Is The Future Of Coding? Warp’s Vision

Summary

In this episode of the Turing Post Inference show, host Ksenia Se interviews Zach Lloyd, founder of Warp, about the future of software development and a new category he’s coined: the Agentic Development Environment (ADE). Zach explains that development is shifting from “development by hand” to “development by prompt,” where developers interact with their tools through natural language rather than typing code directly. Warp positions itself as a workbench built for this agent-first workflow, offering a terminal-based experience that supports the full software development lifecycle.

The conversation explores how Warp differentiates itself from tools like Cursor and Claude Code. While Cursor excels at AI-based autocompletion, Warp focuses on providing a superior agent experience with tight iteration loops, diff review capabilities, and multi-model support. Zach argues that for professional developers, “vibe coding” (letting agents work autonomously without review) is the fastest way to ship buggy software to production. Instead, Warp emphasizes the importance of pairing with agents, reviewing their code, and guiding them carefully.

Zach shares both his excitement and concerns about the AI-driven future of development. He’s excited about democratizing software creation and enabling non-developers to build useful applications. However, he worries about an explosion of low-quality software and emphasizes that agents currently lack understanding of user problems - making human oversight essential. He predicts that within 6-12 months, the agentic workflow will become the primary development workflow, with autocomplete taking a back seat.

Highlights

”Vibe Coding Is the Fastest Way to Ship Crap to Production”

yt-dlp --download-sections "*6:41-7:40" "https://www.youtube.com/watch?v=kz1hATcz-q8" --force-keyframes-at-cuts --merge-output-format mp4 -o "kz1hATcz-q8-6m41s.mp4"

“For pro developers, vibe coding is kind of the fastest way to ship crap to production.” — Zach Lloyd, 6:57

”Agents Will Delete Tests to Make Them Pass”

yt-dlp --download-sections "*22:35-23:30" "https://www.youtube.com/watch?v=kz1hATcz-q8" --force-keyframes-at-cuts --merge-output-format mp4 -o "kz1hATcz-q8-22m35s.mp4"

“If you just give it a goal you got to be really careful. It’ll just start like deleting tests and be like oh the test passed now. It’s like well no not really.” — Zach Lloyd, 22:51

”Become a Senior Developer As Fast As You Can”

yt-dlp --download-sections "*14:09-15:05" "https://www.youtube.com/watch?v=kz1hATcz-q8" --force-keyframes-at-cuts --merge-output-format mp4 -o "kz1hATcz-q8-14m09s.mp4"

“A junior developer right now I think the thing you need to do if you want to stay in software engineering is become a senior developer as fast as you can and learn how to leverage these tools.” — Zach Lloyd, 14:16

”Agentic Workflow Will Be Primary Within 6-12 Months”

yt-dlp --download-sections "*26:03-27:00" "https://www.youtube.com/watch?v=kz1hATcz-q8" --force-keyframes-at-cuts --merge-output-format mp4 -o "kz1hATcz-q8-26m03s.mp4"

“I bet you within like 6 to 12 months that the agentic workflow is primary workflow and the autocomplete workflow is no longer the primary workflow. I put a lot of money on that.” — Zach Lloyd, 26:03

”Development Is Switching From Development by Hand to Development by Prompt”

yt-dlp --download-sections "*0:55-1:55" "https://www.youtube.com/watch?v=kz1hATcz-q8" --force-keyframes-at-cuts --merge-output-format mp4 -o "kz1hATcz-q8-0m55s.mp4"

“Development right now is switching from development by hand to development by prompt.” — Zach Lloyd, 0:55

Key Points

• Agentic Development Environment (0:43) - Zach coined the term ADE to describe a workbench built for agent-first development workflows, distinct from traditional IDEs
• Development by Prompt (0:55) - Development is shifting from typing code to prompting agents in natural language
• Terminal vs IDE (1:39) - Warp looks more like a terminal than an IDE because it supports direct input and launching agents through English
• Warp vs Cursor (3:55) - Warp handles the full software development lifecycle while Cursor excels at autocompletion; Warp is the #1 agent on Terminal Bench and top 5 on SWE Bench
• Warp vs Claude Code (5:22) - Warp is multi-model and offers GUI-based diff viewing, while CLI tools like Claude Code lack file editing capabilities
• Vibe Coding Concerns (6:57) - For professional developers, vibe coding is “the fastest way to ship crap to production”
• Stack Overflow Survey Findings (7:04) - Top complaints about agentic tools: code with subtle bugs and code that’s hard to comprehend
• Warp as Meta Tool (7:53) - Warp plus Cursor was identified as the best combination in Focus AI’s coding agent testing
• Single Workbench Vision (9:23) - Warp’s vision is to be a single integrated workbench rather than requiring multiple tools
• Context Sharing (10:28) - Warp can share context across terminal panes, using server output in one pane as input for an agent in another
• Junior Developer Level (12:29) - Warp can build at junior developer quality and solve problems, but needs guidance from experienced developers
• 5-10x Productivity (13:43) - Senior engineers embracing agentic workflows can achieve 5-10x productivity gains
• Learning Accelerant (14:36) - AI tools can be incredible learning tools for understanding codebases and explaining code line by line
• Why Rust (15:38) - Warp uses Rust for performance, safety, crash prevention, and cross-platform native compilation
• Electron Prototype (17:17) - Warp prototyped in Electron first but hit performance issues, leading to the Rust decision
• Server in Go (18:11) - Warp’s server is built in Go; if starting over, they would use Rust on the server too
• Python for ML (18:59) - Python dominates ML because it’s ergonomic, has great data libraries, and is interactive for notebook-based work
• Democratizing Software (19:49) - AI lowers barriers to building software; Warp’s head recruiter built an MCP server connecting to their ATS
• Fear of Bad Software (20:49) - Concern about explosion of low-quality software that doesn’t actually solve user problems
• Agents Delete Tests (22:51) - Without careful guidance, agents may delete tests to make them pass rather than fixing actual issues
• 6-12 Month Prediction (26:03) - Agentic workflow will become primary within 6-12 months, replacing autocomplete as the main workflow
• Code on Warp Launch (28:51) - Major launch featuring diff review with agents, Sublime-level file editing, and slash commands like /init and /codereview

Mentions

Companies

• Warp (0:32) - Agentic Development Environment founded by Zach Lloyd
• Cursor (3:51) - Popular IDE known for AI-based autocompletion
• Anthropic/Claude (5:22) - Claude Code is a CLI-based terminal agent tool
• Google (5:41) - Gemini CLI mentioned as a single-provider tool; also has “Jules” for autonomous coding
• OpenAI (26:31) - Has Codex for autonomous agent workflows
• Cognition (26:25) - Company working on autonomous coding agents
• Microsoft (4:26) - IDE compared to Microsoft Word as file-editing first interface
• Salesforce (15:57) - Example of a product where users don’t care about tech stack
• Zed (3:28) - Editor whose founder Nathan mentioned needing an agentic development environment
• Stack Overflow (7:04) - Survey data on developer complaints about AI tools
• Focus AI (7:58) - Did testing report on coding agents where Warp performed strongly

Products & Technologies

• Claude Code (5:22) - Anthropic’s CLI-based terminal agent tool
• Gemini CLI (5:41) - Google’s command-line AI tool
• Codex CLI (5:43) - OpenAI’s command-line tool
• Terminal Bench (5:11) - Benchmark where Warp is the #1 agent
• SWE Bench (5:14) - Benchmark where Warp is a top 5 agent
• Rust (15:33) - Programming language Warp is built in for performance and cross-platform support
• Go (17:59) - Language Warp’s server is built in
• Python (17:37) - Dominant language for machine learning work
• Electron (16:37) - Web-based framework; Warp prototyped in Electron before switching to Rust
• Mojo (17:44) - Mentioned as a potential newer language contender
• MCP Server (20:33) - Warp’s head recruiter built one connecting to their applicant tracking system
• Jupyter (19:04) - Notebook environment where a lot of AI work starts
• Docker (9:07) - Container tool that’s easier to use through Warp than an IDE
• Terraform (9:08) - Infrastructure tool mentioned as easier through Warp
• Sublime Text (29:29) - Warp’s file editor now at Sublime Text level of richness

People

• Zach Lloyd (0:22) - Founder of Warp, has philosophy of science background
• Nathan (3:28) - Zed founder who mentioned the need for an agentic development environment
• David Deutsch (27:27) - Physicist and quantum computing researcher, author of “The Beginning of Infinity”
• Douglas Hofstadter (28:30) - Author of “Godel, Escher, Bach” mentioned as similar worldview book

Surprising Quotes

“For pro developers, vibe coding is kind of the fastest way to ship crap to production.” — 6:57

“You need to change system prompt and word like don’t build shitty software. Just don’t. No shitty software please.” — 22:35

“If you just give it a goal you got to be really careful. It’ll just start like deleting tests and be like oh the test passed now. It’s like well no not really.” — 22:51

“A junior developer right now I think the thing you need to do if you want to stay in software engineering is become a senior developer as fast as you can and learn how to leverage these tools.” — 14:16

“I bet you within like 6 to 12 months that the agentic workflow is primary workflow and the autocomplete workflow is no longer the primary workflow. I put a lot of money on that.” — 26:03

Transcript

0:00 No shitty software, please. Development right now is switching from development by hand to development by prompt. For pro developers, like vibe coding is kind of the fastest way to ship crap to production. Holy I can just like ask the computer to start making my…

0:22 Hi, Zach and welcome to the Inference show by Turing Post. I’m very glad to have you here. Yeah, I’m excited to be here. Thanks for having me as guest.

0:32 I’ve been actually playing with Warp and it’s mind-blowing how it builds applications just doing everything by itself and I wanted to ask you, you call it an agentic development environment. If you can tell what it is in relation to terminal to CLI to IDE, what is an agentic development environment?

0:55 Yeah. So the basic idea is that development right now is switching from development by hand to development by prompt. And so what I mean by that is like you know in most of my career if I wanted to build something I’d open up IDE and I’d start typing some code or I’d then open the terminal and type some commands. And what I do today is I open Warp and I type in English or natural language or I literally just talk to Warp and ask the computer to start doing something for me. And the idea is that as the developer workflow shifts to be more agent first, what you really want is a workbench that is built to support that workflow.

1:33 And so like what should that look like? That’s like what we’re figuring out at Warp. I think it looks a lot more like a terminal than it does like an IDE because what you want is this direct input and the terminal the way that it’s always been set up is like you type commands you get answers. Today what you do is you speak English and it launches agents. And so Warp began as a terminal but now the primary way of interfacing with it is through natural language. It no longer makes sense to just like call it a terminal. It does much more than a terminal. It will write your code. It will give you the ability to pair and partner with agents to improve the code by looking at the diffs. It gives you some file editing capabilities.

2:18 So some things that are typically in the IDE are now in Warp. And so we just thought like hey none of the existing labels for these tools actually fit and like when you sort of like stick to the existing categories you run the risk of not building the thing that’s actually the best for the new workflow. So that’s how we’re thinking about it.

2:34 Did you come up with this term? I did. Yeah, I did come up with it. We talked about a lot of different ways of naming it. We were like, you know what, it’s similar to the IDE. We liked how it kind of sounded like the IDE but switch the I with an A and call it an agentic development environment. And I think it’s catching on. Like I’ve heard other people now say that that’s what the world needs and that’s how developments can be done which is pretty cool. So it’d be fun to have actually coined a thing here. We’ll see.

3:01 I was actually going to ask is other companies who already call themselves ADE? I think that we’re the only ones I’ve seen where it’s like the agent… we don’t even call ourselves an ADE like right now if you go to our homepage we call it the agentic development environment. So it’s a little bit of a kind of proper noun at this point. I don’t know. I was just watching another podcast live stream with a couple other engineers who I respect and at the end of it one of them I think it was the Zed founder Nathan was like you know what the world needs is like an agentic development environment. So I do think it’s catching on.

3:36 They’re just logical after IDE. If you think about the current environment with coding agents, how Warp stands out - like what actually defines you and differs you from let’s say Cursor which is super popular and I don’t know Claude Code which everyone uses.

3:55 So compared to IDEs like Cursor - Cursor is the most popular - I guess there’s two things. So one is like because we’re more at the terminal level, it’s like a wider range of tasks that you would want to do with Warp. So you know, Warp is really good for like the full software development life cycle. So from setting up projects to running a coding agent to then deploying the code that’s been built by the agent to then debugging production issues. You wouldn’t really want to do that from the IDE in general because it kind of looks like Microsoft Word. It’s really a file editing first interface. And those are all tasks that you would want to do from something that looks a little bit more like a terminal.

4:36 Warp, I would say also like Cursor, the real sweet spot for Cursor, the thing that they really nailed which is very very valuable is autocompletion. So like AI based autocompletion. So you type, you see the ghost text. I think Warp’s agent, like the actual agent experience in Warp is superior to the agent experience in an IDE right now. Like the agent experience in the IDE tends to be via a chat panel and because it’s like a UI that’s off to the side there’s like a weird… it just is not as good of an experience. And the actual quality of the agent in Warp is better. Like I don’t think Cursor publishes the benchmarks of their agents but Warp is like the number one agent on Terminal Bench, it’s a top five agent on SWE Bench. So I know for a fact that the quality of Warp’s agent is validated against the top benchmarks.

5:22 Versus Claude Code it’s more interesting. So Claude Code works quite similarly to Warp in the sense that it’s based in the terminal and the primary thing that you do in Claude Code is like you prompt it and then the agent goes and does something for you. I think that there’s a couple advantages that Warp has compared to it. So one is like Warp is multi-model. So like if you use Claude Code or you use Gemini CLI or Codex CLI it’s like you’re locked into using one provider’s models whereas with Warp sometimes we’ll use a mixture of models. We’ll always give you like the most state-of-the-art model.

5:52 The other thing is like these CLI based tools have no ability for file editing and so it’s not easy to view the diff or really tightly iterate as they create code and so they tend to be like awesome vibe coding tools is what I would say. So like you know you can run Claude - people who are using Claude very heavily might be like running like five Claudes building different parts of their marketing site or whatever. But if you’re like a pro developer the way to have more success using these agentic tools is to carefully pair with the agent and see at every step what the agent’s doing and have the tools to guide the agent by literally doing a code review of the agent’s code and redirecting it and seeing its diffs making sure it’s sticking to the plan.

6:41 And Warp is just a much better interface for that because we have the full GUI experience whereas if you’re like Claude Code you’re limited to just text-based output of a terminal app of a CLI app. And so Warp is really trying to build out this like… it’s not that we’re against vibe coding it’s just that for pro developers, vibe coding is kind of the fastest way to ship crap to production.

6:59 Right, you vibe code something and then you forget about it. Yeah, or if you look at the Stack Overflow data from the most recent survey it’s like the number one and two complaints of developers using these agentic tools is like the tools generate a lot of code that have subtle bugs. The code is hard to comprehend. It takes sometimes more time to understand what the agent is doing than if you had written it yourself.

7:25 And so, you know, Warp - our view is like build something that is very very agentic but also can be used for real production code bases. And to make that work well, it’s all about can you comprehend what the agent’s doing? Can you guide the agent carefully? Can you correct the agent when it does the wrong thing? Can you make the agent sort of work according to your team’s coding conventions? And so I think we just have this intrinsic advantage when it comes to tight pairing with the agent that these other CLI based tools don’t have.

7:53 When I was playing with Warp, I thought about it as a meta tool. We recently did this report with the Focus AI about - we just tested different coding agents and Warp came out so strong that my husband who was running the testing asked me if you can hire him as an evangelist. So… Awesome.

8:14 Yeah. But his main recommendation was Cursor plus Warp as like the best combination together. So I think at this moment it’s more like a combination of tools. For me it felt a lot like a meta tool because I’m more of an amateur builder and developer. I’m not a professional developer. So for me Warp helps do all the steps that I might not know. And my question is about the vision. How do you see Warp in development? Is it alongside the other tools or will it become an actual meta tool for professional developers as well? How do you see this development?

8:48 Yeah. So I think today we see people using it in all sorts of different ways. Like it’s very common to use Warp plus an IDE. And so you could have Warp do all the things I mentioned before like the project setup, you’re having dependency issues, you want to do all your git stuff through natural language, you want to do all your Docker work or Terraform - all these things that are just unpleasant to do in an IDE. You could just do in Warp. Then if you want you can go into the IDE and do the coding more through that. That’s totally fine. People use Warp that way.

9:20 Increasingly I don’t - that’s not the vision. The vision is more like you can just use Warp as your single workbench. You know the reason I think that’s kind of where things are going to go is because you’re going to have less and less need over time to do a ton of autocompletion and hand file editing. And really the type of file editing that you’re going to need to do is like it’s all about adjusting the path of the agent and giving feedback to an agent as the agent’s working. And that’s what we’re building for in Warp.

9:48 But it’s totally very common for people to use Warp plus Cursor these days. And so that’s a cool workflow. We also have the other workflow of like people use Claude Code inside Warp or use like… and that works too. You know developers you can say one thing about them - it’s like they like picking their tools and they’re best at optimizing their own workflow. And so what some people will do with Claude Code is very similar. They’ll let Warp do all of stuff that they think of as terminalish stuff and then they’ll run a Claude Code session to do the coding.

10:16 There’s no need to do that. Like if you want to pay for two subscriptions or whatever, you can do that. But our vision is that there’s a lot of value in having a single integrated agent experience. For instance, if you do the coding directly through Warp rather than doing it by running a CLI app, Warp can share that context across all of your terminal panes. And like you can when you’re running a server in one pane, use the output of that server as input for an agent that’s running in another pane.

10:45 And so there’s all this awesome value that you get from an integrated agent experience that we think people will eventually… a lot of people are doing it this way, but for people who prefer using multiple tools, we think they’ll eventually be like, hey, it is actually easier to not do the context switching. But it’s totally up to developers and developers are smart and want to pick what’s going to make them most productive.

11:06 It’s a lot of rethinking of workflows for developers right now. And what do you think would disappear from today’s workflow once that type of meta tools exist?

11:17 I think that the existing tools will just become less used. Like I think you’re always or for a long time you’re going to want to be able to use a file editor because there will just be cases where it’s either faster for you to do it yourself or you started with an agent, it’s not doing it right and you want to get in there and change it or it’s like you’re working in a language that the agent isn’t good at. There’s all sorts of reasons that you’ll want to do that for the same reason that sometimes as a very rarely sometimes developer you need to look at the machine code or something.

11:47 So it’s like there will be an abstraction that’s more convenient to use than the existing tools and you’ll use that for like 95 or 98% of the time but when you have to get closer to the metal so to speak you’ll want to go into the code editor and I think that tool will exist. So I think it’s more like the predominant way of working is going to switch to be agent first but you’ll have these fallbacks for when you want to go and hand tweak stuff. That’s my guess.

12:17 I would say now Warp can be a junior developer already. You’re saying that we will be agent first. So does it mean that this tool can be like a senior developer or completely independent developer? How do you see it?

12:29 Junior developer analogy is pretty good actually in that if you give Warp a spec of things it should do or you iterate with it on a plan and are like go build this, it can build it at the same quality as a junior developer sometimes even better. It’s pretty crazy. And not only can it build things it can generally solve problems which I think is an important distinction. So it’s not just a coding tool, say like solve my dependencies tool or figure out my server’s crashing tool. It can do a lot of the different tasks that a developer can do, but it needs to be guided right now by a more experienced developer.

13:04 The way I see it evolving is like the actual types of tasks that it will be able to do will become more and more complex over time and it will be able to do things that good senior developers can do. And this will be driven by improvement at the model layer, but also improvement at the app layer. Improvement in terms of the context that’s being provided, improvement to how the tool is learning your team’s knowledge and your organization’s knowledge so that it’s not having to relearn everything from scratch every time it starts.

13:35 But I still think over the next few years at least you will have senior engineers who fully embrace this way of working who find that they can do like 5 to 10x the amount of stuff and a lot of the job becomes how do I effectively communicate what needs to be done and how it should be built so that an agent can do most of the brunt work of doing it and then I will review it and make sure it’s right and isn’t going to introduce bugs or security holes and it fits into our overall architecture and the product road map and all that stuff.

14:09 So I think that the job will evolve so that it’s more like that. But a junior developer right now I think the thing you need to do if you want to stay in software engineering is become a senior developer as fast as you can and learn how to leverage these tools. Yeah.

14:21 How can the tools help you with that? Yeah. So I think that that’s actually an overlooked thing. You know, there’s a lot of fear around the tools replacing, but the tools can also be a huge accelerant to learning. They can be tools for that.

14:36 They’re incredible learning tools if you use them that way. And so something like Warp is awesome - not just doing stuff, but explaining. It can be a very powerful shortcut to learning a new codebase for instance. Or it can walk you through code line by line, explain what’s it doing, why is it working this way, or it can suggest multiple ways of potentially building something and give you the pros and cons of going with each approach.

15:04 It’s very similar to ChatGPT in the sense of you can use ChatGPT to cheat on all your tests and write all your essays if you want and that’s a choice that you’re making or you can use ChatGPT as a way to accelerate your learning for some topic. It’s a little bit up to the intent of the person who’s using it what they want to get out of it. And so it’s a phenomenal learning tool if that’s your intent going in.

15:27 I have two questions about languages. You chose Rust for Warp’s foundation. And so what does Rust enable here that wouldn’t have been possible otherwise?

15:38 We chose Rust because we wanted to build an app that felt fast. Honestly, first of all, developers I think are a uniquely sensitive group of people to the performance of their apps. They’re also sensitive and interested in how apps are built, which is just not a thing for any other consumer segment. I don’t think anyone who’s using Salesforce is like, is this built using React or whatever. But developers actually care about this. They’re curious about it.

16:04 So we chose Rust. The primary thing was performance. Secondary thing was or probably of equal importance was safety and just not having a lot of crashes. I really strongly prefer strongly typed languages. Rust makes it a little bit harder to write a slow app when it comes to the way the memory management works.

16:24 And then the other great thing about Rust is like if you want to write a cross-platform desktop app which is what Warp is - it’s a native fully native app on Mac, Windows, Linux and also want to support the web - I’d say you really have two choices today. You can use the web platform and do it in an Electron app which is what most people do or you can use Rust and compile it natively for all these other platforms.

16:47 Rust is definitely the harder path and I don’t know that I would… I would only do it if you have an app that has these types of performance characteristics where it matters and it does matter for Warp. Warp is streaming through gigabytes of data of textual data and it’s very annoying if it all of a sudden starts to stutter or there’s all these problems with it. Rust gives you really nice primitives for doing asynchronous work.

17:09 We just end up thinking for Warp it would be better to go one level lower and use something that’s closer to the metal than using Electron. And we did prototype Warp in Electron first and we started to run into some performance issues pretty quickly and we’re just like let’s pick the thing that we think is going to result in a better product, even if it’s a little bit harder to build.

17:30 Now when you work so close with AI and AI is famously - machine learning is famously using Python all the time. How do you feel about it? And do you think that the ADE era will shift parts of the stack toward Rust or newer contenders like Mojo or other languages or will it still be Python?

17:51 Great question. So most of our agent logic is on our server which is not built in Rust. The server is built in Go. Again, sort of the same question applies - why not Python? If we had to do it again actually, and we talk about this internally, we would probably use Rust on the server as well because there’s some code that we want to be able to share between client and server. Go is okay. I’m not nearly as big of a fan of Go as I am of Rust, if I’m being honest.

18:18 Go is kind of in no man’s land. I don’t think we’re getting that much from it compared to either using Rust or using Python. Like Python would be nicer I think for doing the more AI oriented work. But the truth is you can build anything in any language. We’re able to build and iterate fast enough on the server side with Go that it hasn’t been an issue.

18:37 Why do you think they stuck with Python for machine learning? I think it’s the main language because it’s very very easy to use and it’s very ergonomic. Also, for whatever reason the core data libraries like the stuff that’s done at the data science level - the Python libraries have caught on for that. And also it’s a very interactive dynamic language.

19:00 I think a lot of AI work starts in a notebook, it starts in Jupyter and so you want something that’s not compiled that’s super interactive that is very high level also. So it’s not hiding any of the concepts. It just is making the concepts of doing the machine learning front and center. So I think for those reasons it makes sense to me that it’s the primary place where all this ML and LLM work starts.

19:29 Now when you’re not just in software development anymore, you’re in an agent development world. What excites you the most and what are your concerns about the future that you help build?

19:42 The thing that excites me the most - and I’ve been a developer for a really really long time, I don’t want to go into how long - but the thing that excites me the most is just how much easier it is to build awesome software and how much more is possible to be built by fewer people and also how it expands the number of people who can build software.

20:05 So it gets rid of a lot of the gatekeeping around who gets to build software and apps. I think that historically that’s been a real problem and there’s relatively few people in our society who know how to build software and software is such an important thing in the world. So it definitely democratizes that, it lowers the barrier.

20:25 We have people on our team who are not software developers who are making apps to help themselves do their own jobs better. Our head engineering recruiter has built an MCP server that connects to our applicant tracking system and he’s not a developer and it’s so cool that he now has this ability to make his own job so much easier. So that’s what excites me.

20:49 What concerns me is that there’s potential to be a lot of really crappy software that’s built and brought into the world. Not just a builder of software but as a user of software - obviously I hate using shitty apps. It’s a real… I hate it. I hate when I go to like my pharmacy’s website and I just want to refill a prescription and it doesn’t work. I can’t log in. It doesn’t find my stuff. The buttons don’t work.

21:16 And so I’m just afraid as a user that this is going to lead to an explosion of really shitty software that I have to endure using. And it’s one of the things I fundamentally don’t get actually about vibe coding is like it’s awesome that everyone can build software, but what matters is actually not just the pure volume of software that goes out in the world. It’s does this software solve problems for people? Does it solve them in a useful and fun to use way?

21:44 And my worry is that there’s going to be just a lot of bad shitty software shipped and that even when you look at the more pro use cases for the agentic stuff, my worry is that it’s going to lead to security holes or unmaintainable code. There’s good reasons that the practice of software engineering has evolved the way that it has.

22:05 That’s why I think that in order for this stuff to really work in production, you need to go way beyond vibe coding to building a tool set that lets you really carefully specify what you want to build. But not just that, that lets you have a very short tight feedback loop and make sure that you understand what’s being built and are guiding the agents to do it the right way. I think that’s where we’re at today. And so my fear would be there’s just going to be a lot of crappy software.

22:35 You need to change system prompt and word like don’t build shitty software. Just don’t. Yeah. Like that. Like if we could just tell it that maybe that would just solve it. No shitty software please.

22:44 It’s like a kind of savant. It will do things where if you just give it a goal you got to be really careful. It’ll just start deleting tests and be like oh the test passed now. It’s like well no not really. And so it doesn’t understand the user problem that it’s solving which is to me the absolute key thing if you want to even be a good engineer as a person - you always have to be like why am I building what I’m building, how is it useful to someone. And until the agent can understand the user problem I think it’s bound to produce stuff that’s not that awesome. You need a human to be providing that context.

23:19 Right, human in the loop. Do you think at all about AGI and super intelligence? I do. I’m not qualified like prognosticating on when it’s coming. I don’t know. All I’ll say is the stuff is changing super duper quickly and where we’re at now compared to 6 months ago implies a sort of exponential curve. I don’t have enough knowledge to say how long will that exponential continue? When will it become more of an S-curve?

23:46 I do think the typical pattern is you’ll have some period of exponential and it will become an S-curve. It could be that for AI though it becomes sort of recursively self-improving. I think this is what the people who are on the ASI side think is going to happen. I don’t know. It’s not there yet, but the rate of progress is obviously super fast.

24:04 What I heard from you is that human still is needed. It’s needed in this loop. The bet that we’re making I guess in relation to this is we’re betting that the models will improve and that they will improve rapidly. I do believe that these agents will become more and more autonomous and be able to do harder and harder tasks and that you need to build - we’re building Warp for that future.

24:25 However, if you look at where we’re at in terms of the adoption of the agentic workflow and the current state of the models, the first big AI breakthrough was autocomplete with Cursor. I think we’re starting to see the widespread adoption of agents, but most developers aren’t using agents yet. That’s just a fact. It’s a foreign workflow that developers need to adopt.

24:43 And then I think the third thing that will happen is that the agents will become more and more autonomous. Meaning you might have system triggers - hey, there was a crash, an agent then goes and investigates. So I think that’s coming. What’s needed today is this tight iteration loop. And even in the more autonomous future, what’s needed is maybe it’s not an iteration loop, but it’s an orchestration loop. It’s some way of expressing intent and managing what the agents are doing even as you have more of these agents at scale.

25:14 So in terms of time frame for that, again, I think this stuff’s happening very quick - like 6 to 12 months. I think it’s going to look pretty different. But today as people are just getting used to building with agents, I do think you need that really tight loop in order to ship production quality code.

25:28 It’s very hard to make predictions. But if we talk about adoption and some time frame for that for software developers, what do you think will happen in the next let’s say 12 months?

25:39 Where I think we’re at right now is the mainstream developer has adopted an autocomplete tool like Cursor. In terms of agentic tooling adoption, what you see is AI enthusiasts plus non-technical people like PMs, designers, and an increasing number but still relatively small number of bread-and-butter pro developers being like, I’m going to start every task with a prompt.

26:03 I bet you within 6 to 12 months that the agentic workflow is the primary workflow and the autocomplete workflow is no longer the primary workflow. I put a lot of money on that. That’s the bet that Warp is making. And so I think that adoption is going to happen really quickly as developers fully get like holy… I can just ask the computer to start making my change, provide it the right context, tell it how I want it to do it and it will do it.

26:26 And then I think the more autonomous workflow which is like you might look at a company like Cognition or I think Google has a thing called Jules or OpenAI has a thing called Codex - that’s also happening right now it’s just happening for a very small slice of tasks that agents can actually do that way. And I think a year from now or 6 to 12 months from now, that will be becoming more and more mainstream as well.

26:53 And so the simpler tasks will start to be fully automated and the real meat of development which is building features, fixing bugs, fixing stuff in production will be this agentic workflow. But I do think the full automation of simpler tasks is going to be happening in parallel as well. So it’s going to be wild. And then like two years from now, I don’t know. I just straight up don’t know. Three years, I really don’t know.

27:17 Crazy fast. And my last question is about a book. If there is a book that you read this year that you can wholeheartedly recommend.

27:25 Yeah. So I’m reading this book called The Beginning of Infinity by David Deutsch who’s this pretty famous physicist who’s also a quantum computing person. It’s a pretty cool book. It’s taken me a long time to get through it because it’s a pretty dense book. He’s a very very smart guy. Have you read this book? Started.

27:44 I think it’s really fascinating. My background even before I did engineering, I have a philosophy of science degree. And so it’s a lot about what makes human intelligence special. He’s like human intelligence and the ability to produce explanations that are hard to vary is a really powerful physical thing in the world. And talks about the nature of progress and what that enables. He talks a bunch about quantum physics and he’s a subscriber to the multiverse theory of the universe.

28:16 And so it’s really one of these books that’s kind of all-encompassing explanation type book which I think Guns, Germs, and Steel is one that I read a long time ago that’s kind of in that camp or Godel, Escher, Bach by Douglas Hofstadter that’s another one in that camp. So it’s one of these really cool I’m going to give you my whole worldview type books. I do recommend it. It will change the way you think about the world a little bit.

28:41 That’s cool. Thank you so much. Thank you for your time and insightful conversation. Cool. I really appreciate you having me on. This was awesome.

28:51 So this week we have a major launch that we call Code on Warp that should be live by the time this episode comes out which has a few key features. So one is the ability to actually review diffs with an agent and pair on editing those diffs directly within Warp. And so the workflow here is like you start a coding task, the agent produces a bunch of diffs and then you basically as a developer get to code review the agent’s diff and have it correct itself or if you want you can edit it.

29:20 The second thing is we’ve invested a bunch in file editing capabilities in Warp. And so you’ll find the actual file editor is a much richer experience kind of at the level of Sublime Text now in Warp where it’s not meant to be a place where you start by typing code but it is a place where it’s great to iterate and edit an agent’s code.

29:39 And then the third thing is that the agent now has the ability to spin up a new project or bring one down from GitHub and create project-based rules and a project based codebase index. And we’ve made that whole flow a lot simpler for people. So including the ability to have different slash commands which you’re familiar with from other agentic tools - you’ll find similar in Warp like /init or /codereview and so these things that if you want to make Warp your primary coding tool the experience has gotten way way way better this week.