Meet The DevTool Founders Building For AI Agents

1. 0:00 – 3:00

   Meet the Founders

   1. SPSpeaker0:00

      Code doesn't matter anymore. What's gonna matter and what's left is context, is understanding problems deeply, building relationships with customers.

   2. SPSpeaker0:08

      If you're not accustomed to change, you should get used to change.

   3. SPSpeaker0:11

      The AI agents have taken so much toil out of building software that now we can get a lot better stuff out a lot faster, a lot higher quality. It's going to make, uh, engineers more valuable, not less valuable.

   4. SPSpeaker0:26

      [upbeat music] Infisical is an open source identity security platform. We're working with companies like Nvidia, LG, Hugging Face to help them manage their secret certificates and other sensitive resources across infrastructure.

   5. SPSpeaker0:39

      We make it easy to get up and running with agents and assistants using open models.

   6. SPSpeaker0:43

      Resend is the email API for developers. Uh, you can think of it as the Stripe of email.

   7. SPSpeaker0:48

      Recall.ai is the API for conversation data. We help developers record, store, and understand data from video conferences, phone calls, and in-person meetings.

   8. SPSpeaker1:00

      Greptile is building agents that review and test pull requests with full context of the code base. We do this for companies like Nvidia and Coinbase, where we review more than three billion lines of code a month and flag anti-patterns and bugs.

   9. SPSpeaker1:11

      Firecrawl is a web data API for AI agents and developers.

   10. SPSpeaker1:15

       Porter is a platform that manages compute for startups and any cloud provider.

   11. SPSpeaker1:19

       We've create AI-powered intelligent documentation for enterprises and startups.

   12. SPSpeaker1:25

       We provide, like, local LLMs, and we, you know, make it much easier for people to train models.

   13. SPSpeaker1:30

       We make in-app subscriptions easy, help developers make more money.

   14. SPSpeaker1:36

       Our main audience today is developers using coding agents.

   15. SPSpeaker1:39

       The main audience is developers, but increasingly more so it's becoming agents.

   16. SPSpeaker1:43

       Developers. Uh, we definitely tailor to agents, and we try to create the best experiences possible for them. But we see agents as distribution and developers as the users.

   17. SPSpeaker1:55

       The customer is usually agents, so we spend a lot of time building interfaces for agents, for instance, CLIs and skills and so on.

   18. SPSpeaker2:01

       We think of our audience as anyone who wants to run high-scale infrastructure, so that could be developers, and increasingly it's also agents.

   19. SPSpeaker2:10

       A-at, at time, like even for example with OpenClaus, Infisical sort of like becoming the infrastructure layer for OpenClaus to consume different credentials and communicate with each other.

   20. SPSpeaker2:19

       Oh, it changed completely. From an organization standpoint, uh, we just started to create one OpenClau agent after the other internally. So we started with a marketing one, then we added a sales one, then we have one for infrastructure, one for DevOps. Uh, so it really influenced both our product and our company.

   21. SPSpeaker2:37

       We're definitely making all of our products more accessible to agents, which means making everything into a CLI, making data more exportable so that you can use it with Claude Code.

   22. SPSpeaker2:49

       It makes our output much faster. Previously, you would do, like, one feature, but now you can do, like, 10 features all in parallel in one go.

2. 3:00 – 4:22

   Building for Agents First

   1. SPSpeaker3:00

      If I were starting the company today, this is something that we've already done, but I would just do it earlier, uh, which is consider agent experience as a first-class, uh, part of the interface.

   2. SPSpeaker3:11

      One thing we'd do differently is redesign the whole product experience so it's accessible to agents. We're working on that right now. Whether that's Ollama's documentation or it's other capabilities that we offer to humans and developers, uh, making sure that they're immediately accessible to coding agents as well.

   3. SPSpeaker3:25

      If I were to start the company today, probably get much more comfortable with deleting things that we know are not going to be useful after some period of time. I think emotionally detaching oneself from the product because you know that you'll build something today, and it'll be useful for a couple of months, and then something drastic will change about the world, and that part of the product will be useless. I think having the courage to ruthlessly delete that which is no longer useful, it would become increasingly very important. I would build that muscle earlier.

   4. SPSpeaker3:48

      Yeah, I think we would maybe just go all in on some of the agentic infrastructure, things like sandboxes. We care about and are doing those things now, but there is definitely an advantage to having that a part of your core DNA from the very beginning.

   5. SPSpeaker3:59

      The biggest difference would be to use agents more [chuckles] , um, to use Codex, use Claude Code more.

   6. SPSpeaker4:05

      I would be 100 times more ambitious. The capabilities that models give you today make the impossible possible, and not only that, they make it table stakes. All of a sudden, because the bar has been raised, you have to do whatever you can to get there.

3. 4:22 – 7:15

   Biggest Early Mistakes

   1. SPSpeaker4:22

      Oh, boy.

   2. SPSpeaker4:23

      Nothing.

   3. SPSpeaker4:24

      [laughing] I think any time we started to drift away from if not the, like, exact persona of developers, but the spirit of developers, right? Of getting something done quickly, of being very direct, of being very clear with your pricing. All of these things that developers value in an experience is actually just good user experience. And, you know, through the various iterations of sa- you know, go to market and things like this, we've, we've sometimes deviated away from that, and I think, you know, with maybe some small exceptions, those have been mistakes.

   4. SPSpeaker4:58

      I think another mistake we made was that at the beginning we were extremely opinionated at, about how our APIs and how our, our tools should be used, and sometimes you try to make it a little bit more generic, and that creates even more conflict. If you're an authority in the field, you should be very opinionated.

   5. SPSpeaker5:13

      I think in the early days, our understanding of developer experience was a lot about, hey, we gotta have the perfect UI and the perfect design, which is not actually what developer experience is. And, and right now, you know, a lot of what we think developer experience is, is how fast can they get to the right, you know, aha moments and the, the functionality that they need to use as part of their environments. And so we are much more optimized for the speed and efficiency and, and also how clear it is to use Infisical.

   6. SPSpeaker5:41

      For a long time, we didn't trust our own instincts strongly enough, and that was partly because as new college graduates that hadn't spent a lot of time in large companies, while we had, in some regards, the same job title as many of our users, which were developers, uh, we sort of saw that we weren't really like them. They were working in these large companies, and their day-to-day was very different. Their way of approaching these problems was different. And I think we went a little too far in that direction, where we leaned far too heavily on understanding our users, which was important, but not trusting our own instincts for what a good product looks like. When we started to realize this, we realized, wow, there's a lot of aspects of our product that have become sort of full of cruft, and if we had just trusted our instincts more strongly, we would've ended up with a product that was just more pleasurable to use. Because there's, of course-a level of understanding you can get from using the product yourself and trusting your own judgment on the product that you just can't glean as easily from talking to a person

   7. SPSpeaker6:34

      One of the things that we did early on, which we've actually changed very recently, is the way we think about documentation. In the old world of documentation, uh, you would have a lot of exposition. Uh, you would have it spread across multiple pages. Uh, you'd have embedded tutorials. Um, but with agents, one thing that we've been really conscious of is managing the context window. Now, we've redesigned a large portion of our docs to be able to fit inside a coding agent's context window very easily, uh, so when the agent is working through the integration process with the API, uh, it gets much higher accuracy and it makes much fewer mistakes.

4. 7:15 – 9:22

   Do Founders Still Write Code?

   1. SPSpeaker7:16

      No, I don't write code. I write prompts for humans and agents.

   2. SPSpeaker7:19

      No, our agents do [laughs] .

   3. SPSpeaker7:22

      Yes, that's, that's the right answer.

   4. SPSpeaker7:23

      Oh, yeah, every single day I push a PR, for sure.

   5. SPSpeaker7:26

      Do I write code today? Um, good question.

   6. SPSpeaker7:28

      Every day. Most of it's coded by, uh, a, a coding agent.

   7. SPSpeaker7:31

      I, I basically review the code that's, uh, being made by coding agents, and I get to spend a lot more time with the model builders themselves, the hardware companies, and developers building applications.

   8. SPSpeaker7:43

      I don't write production code anymore, but I write code for fun, and I've found I'm writing even more code than just a couple months ago because the coding agents have taken all the toil out of programming.

   9. SPSpeaker7:53

      The biggest unlock now is that even managers or engineer managers that used to have the manager schedule and they couldn't make themselves be a bottleneck for releasing new features now have the ability to do the people management and the meetings, but also write production code.

   10. SPSpeaker8:07

       And I think that's gonna be the expectation-

   11. SPSpeaker8:08

       Yeah

   12. SPSpeaker8:08

       ... going forward.

   13. SPSpeaker8:09

       There was a period of time in the middle when I'd stopped writing code, especially when we were really focused on go to market, but as we've developed a go to market function with really great leaders, I've been able to spend a lot more of my time writing code. And I'm not contributing too much to the main code base, but there is a small number of internal tools that I maintain, and the main reason I do so is so that I'm essentially constantly using the product at least a couple of times a day. I'm constantly using the new features that we're developing, and I can sort of credibly and legibly talk about what needs to change in the product and what's working well.

   14. SPSpeaker8:37

       So I feel like we still write code, but the total amount of code is much more, and the percentage of the total amount of code is from Cl- Claude and Codex, you know, is like 90%, and then self-written code is like 10%. Um, so it's not like we're not writing less code, it's just that the total amount of code is more.

   15. SPSpeaker8:54

       Yeah, it's funny. I probably stopped writing code for a bit, and have now recently, I think like many founders, had to come back to writing code personally just to kind of rebase my assumptions about where the technology is. You hear so much, you see so much, especially online, but you kind of have to use the tools to build an intuition for it. And so literally in the last few weeks, I have been heavily, heavily trying, you know, everything that's out there.

   16. SPSpeaker9:15

       No [laughs] . I use Claude Code for everything [laughs]

5. 9:22 – 12:09

   Most Unexpected AI Discoveries

   1. SPSpeaker9:22

      .

   2. SPSpeaker9:22

      The one really unexpected thing for us that agents do for us, uh, is they bring customers to us. We looked at the ways customers find out about Recall, and we discovered that all the coding models recommend Recall all the time, and this was super surprising to us.

   3. SPSpeaker9:39

      I was very surprised that, given enough context, agents have remarkably good judgment on what we should work on. What I did do early on was give them access to, for instance, some number of customer support tickets, some amount of data on what pull requests we'd opened recently, and then have them critique my plan for what I think we should do over the next few weeks. And, uh, it was remarkable that they had very clear judgment, and I think part of that has to do with the fact that they just have far larger context windows than I have, and they can read, for instance, thousands of support tickets over the last 90 days and understand deeply what each person in the team is shipping, and also see the world very clearly and be able to look at my set of ideas that I've put forward and say, "Number four just doesn't make a lot of sense based on where we're seeing things go," and, "Wow, there's this thing that a lot of users are asking for and you're just not sort of pricing that into your plan for the next few weeks." I was very surprised at how good they are at critiquing our plans.

   4. SPSpeaker10:33

      One unexpected thing would be, so we actually are doing a lot of internal tooling with agents. Um, one of the tools that we have is this very custom vulnerability detection agent that's able to find and detect any risks or vulnerabilities inside our code before they get merged into production. It's much more, uh, capable than everything that's been done before, and it's also very, very, uh, custom to our code base, so it works really great.

   5. SPSpeaker11:00

      This one surprised me. It wasn't, um, it wasn't my use case, but earlier, uh, this year we learned that, um, there are countries without internet that are getting open models delivered on computers to them, and they're doing this because they have access to all this information. And that was kind of unexpected that a, a model that was trained using a lot of data from the internet is now being used in places where the internet doesn't even, uh, reach. And so that was really exciting and surprising to me, uh, that models were an alternative to that.

   6. SPSpeaker11:29

      I think this is an underappreciated thing, strategy, honestly. Um, and this was a- also possible with older f- form factors even with, you know, chat, but I think, you know, agents have taken this to the next level now that they can call these tools and do deeper research. Um, a lot of people talk about execution of things. I think strategy, for all the reasons I've mentioned, matters more than ever right now, and it's really helpful actually. I don't want an agent setting the strategy of the company for me. I want to come with something, but I can prompt it and have it really push back in different places that I wouldn't have expected. Um, and it's actually helped me reveal gaps in my own thinking, so I think that's like a slightly maybe undervalued aspect of interfacing with agents.

6. 12:09 – 14:38

   What's Underrated Right Now

   1. SPSpeaker12:10

      I think it's the proliferation of open models. It's still much of a secret that open models are quickly catching up and are far more efficient to run.

   2. SPSpeaker12:18

      Yeah, I think I will go back to quality. To me, there's no other way. Uh, creating a product with the best quality you can is just

   3. SPSpeaker12:28

      extremely obvious, extremely important. Getting that in practice and making that happen is way harder than, than you think

   4. SPSpeaker12:36

      Recently, there isn't a clear answer to what the best coding agent is. There was a period of time when people were certain that sort of Cursor was better than Copilot. And as time has gone on, the differences between the agents have become more subjective than objective, and there are cohorts of people within these companies that are completely certain that Claude is the best, or Codex is the best, or Cursor is the best. And a really interesting downstream effect of this is that just like in the pre-AI world we saw different people use different IDEs, I suspect a very similar thing will happen with AI coding agents too. What is also interesting is not only are different engineers using different coding agents, increasingly we ask not startups, but fairly large companies, uh, who is contributing to the code base outside of engineering, and of course, design and product are by now contributing to engineering. But we recently heard about finance people contributing to the pricing or the, the billing logic of the, of the product. Or you saw folks from sort of the customer support org that are spinning up agents to fix bugs that they're seeing that are really obvious. And I thought that was very interesting because I think the perfect coding experience, the perfect interface to express intent for a support person or a design person is very different from a backend engineer or an infrastructure engineer. And I think something that is sort of underrated is I think coding agents are, are not actually a winner-take-all market. There's going to be a very long tail of various different products that are very customized for a certain type of person that really just wants to use a certain type of interface, a certain type of location to spin up their coding agents.

   5. SPSpeaker14:01

      One thing that feels obvious to me is that the market is always going to drive intelligence and capital towards the frontier. And so I think a lot of people are scared right now about, "Oh my gosh, everything I'm building now is gonna be, you know, rebuilt by, by agents." Even if in the future your thing could be, you know, one-shotted, if you could one-shot Stripe, AWS, all these crazy bits of software, just given the path-dependent nature of the world, you can still kind of be this part of the layer cake or the onion right now where you are actually a, a foundational piece of infrastructure that doesn't get rebuilt and maybe has, like, a surprising amount of durability, um, not necessarily in the asymptote, but for much longer than maybe some people expect.

7. 14:38 – 20:29

   Predictions & What's Next

   1. SPSpeaker14:38

      So we're still experimenting with, like, you know, doing in-parallel development of, like, you know, features and new product and stuff like that using these coding agents. I feel like there needs to be more of this. Um, you know, for us, we can only handle, like, eight, you know, in-parallel development. But I feel like, you know, what happens if someone does, like, in parallel 100 features in one go, you know, in parallel? That would be very cool. And so the biggest problem with these agents is, you know, still you need verification, you need a re-review process, um, and that is still very lacking. Um, and so I feel like if someone does something like that, that could be very interesting.

   2. SPSpeaker15:10

      I hope everyone is acknowledging the fact that we haven't 100% figured it out yet, and that they should be still ready for more change, because who knew what was gonna happen over the course of the past year? And I feel like people used to ask questions within the span of, like, five to 10 years into the future, and I feel like you just can't do that now. You just cannot predict that far out. Um, so if you're not accustomed to change, you should get used to change.

   3. SPSpeaker15:39

      Code doesn't matter anymore. What, what's gonna matter and what's left is context, is understanding problems deeply-

   4. SPSpeaker15:45

      Mm

   5. SPSpeaker15:46

      ... is, uh, building relationships with customers. And I don't think there's a way to vibe code around that. I think there's gonna be a lot of tumult, but, like, I think software is going to continue. Like, the, the i- the, the, the tr- people translating computer power into, you know, problem-solving for people, I, I don't think is, is, is gonna stop.

   6. SPSpeaker16:07

      We're on a path to full autonomy, and people struggle with that. They, they lean away from it 'cause it forces them to ask questions about their identity, w-what it means to do work, what it means to be human. But once you accept that truth, you can adapt to that future. You can steer your own path through life and, uh, hopefully the path of the world to somewhere that's better.

   7. SPSpeaker16:30

      Something that I'm really excited about is software getting better, faster, more featureful. Like, I think there's this sort of doomer mindset amongst a bunch of people in technology, which is, okay, AI technology, they think it's gonna replace software engineers, but what we've actually seen at Recall is it's increased the need for software engineers. We're actually hiring more people now, uh, than we planned to hire before.

   8. SPSpeaker16:55

      For me, I think the biggest unlock is looking at the hardware systems that are coming in the future. The data center grade models of today are going to become the personal AI assistants that are going to run locally on your computers tomorrow.

   9. SPSpeaker17:10

      More and more tools will get predominantly chosen by agents, not developers, and, and we're already seeing it. Most of our user base and, and, uh, kinda, like, new customers are coming through Claude or ChatGPT or Gemini or, you know, other types of systems like this.

   10. SPSpeaker17:26

       We are delegating a lot of, um, the work to agents, and what will happen is that agents will now make decisions for us. If you're being recommended by Claude and, and Gemini and ChatGPT, you're in a great place, but if you're not, then it's gonna be very hard for you to, to stand out. So you've gotta really think about not only developer experience, but also agent experience.

   11. SPSpeaker17:51

       I'm very excited for a world where all the developer does is come up with and express some opinion, some idea, some element of taste, and agents figure out how to turn that into industrial code that can be shipped to millions of users, that'll be high uptime, high reliability, high scalability. Towards the end of the year, we should start seeing real programmers at real companies with real products start to experience the first taste of truly idea guy programming.

   12. SPSpeaker18:17

       I'm excited about dev tools where the human doesn't even need to be in the loop at all. So an agent can sign up, can pay, manage access. It can perform all these tasks without even having to tell the human about it. I think that'll create a case where an agent will be, be able to use 10, 20, 100x more, um, compute or any other service than the human would've ever been able to.

   13. SPSpeaker18:39

       Incident resolution, so tasks where production infrastructure is jeopardized, there's downtimeThese are in many ways verifiable, um, and I think that we're basically on the cusp of making AI, at least incident response, a reality this year.

   14. SPSpeaker18:53

       Developers will also start going into other, like, industries as well because being a programmer will be so high leverage for automating everything.

   15. SPSpeaker19:03

       So one of the biggest missing points of open source versus closed source, um, was tooling did not really exist. And so that's why the closed source model apps were always very popular. And so I feel like just this year alone, you know, with OpenClaw, you know, with many other open source projects, it made the tooling for open source much better. And you know, I'm very excited for like, you know, even more of this. Um, and this makes like open source get, you know, get much closer to closed source.

   16. SPSpeaker19:27

       I'm not excited, but what I'm expecting right now is that we hit a ceiling because I was saying that for the last two years and never happened [laughs] so-

   17. SPSpeaker19:35

       We went through a big zero to one change from like humans writing most of the code with some assistance from agents. And honestly, since, you know, the late 2025, early '26, we flipped to the agents actually can do 90% of the coding and the, the power relationship or the, the relationship flipped. I don't think it's gonna flip again. That would be my prediction, but again, I'm with Miguel. I've been predicting stasis and normality for the last two years and been wrong every single time [laughs] so-

   18. SPSpeaker20:02

       People want better software, more reliable, faster, covering more features, and the AI agents have taken so much toil out of building software, uh, that now we can get a lot better stuff out a lot faster, a lot higher quality. It's going to make, uh, engineers more valuable, not less valuable. [upbeat music]

Episode duration: 20:29