From Zapier for Devs to Powering 90% AI Agents

Nicolas Dessaigne 0:00 [upbeat music] I'm here with Matt and Eric from Trigger.dev, who recently announced their Series A from Standard Capital. And today we're gonna spend a bit of time talking about their time in YC, their product, and where they think software is headed. Hi, guys. Welcome.

Matt 0:19 Hey, good to see you. [laughs]

Eric 0:20 Thanks for having us.

Nicolas Dessaigne 0:21 Guys, lots to talk about, but before we get into any of that, maybe just start by describing what Trigger.dev is.

Matt 0:27 Yeah. So it's a way to basically add AI agents to your, your product. Um, so you use our SDK, and you create agents, and you add them to your existing product. Um, and we will, like, execute them for you and run them reliably.

Nicolas Dessaigne 0:42 That makes a lot of sense, but it, it's fairly different from what I remember you pitching when you joined YC in 2023. Uh, maybe let's, let's start by just talking about the first versions of that idea when you applied to YC.

Matt 0:55 Yeah. So we, we actually pivoted pre-YC, but we, we launched the first version of Trigger.dev, I think on, like, February the 1st, so, like, quite early in the batch. Um, and it was an async background jobs framework. So if you wanna run work in the background, um, we would, like, do that for you.

Eric 1:11 Although we pitched it as Zapier for developers, and it was sort of the, the meme at the time was, like, there was three or four different things that were Zapier for developers, and we were sort of the first one to, like, capture that. But, um, yeah, we had a big, uh, Hacker News launch that went really well.

Nicolas Dessaigne 1:25 I remember that.

Eric 1:26 Yeah.

Matt 1:26 Yeah.

Nicolas Dessaigne 1:26 Yes. And, uh, you guys came in with the Zapier for developers.

Matt 1:31 Mm.

Nicolas Dessaigne 1:32 Uh, you got into YC. I remember being, uh, really impressed with your sense of design. Like, it was pretty clear looking at your website, uh, that you were... A, you had a really good sense of what developers would get excited about, which is, I think, why you did so well on Hacker News.

Matt 1:46 Mm.

Nicolas Dessaigne 1:46 But everything just looked, uh, like it was done with care.

Matt 1:50 Mm.

Nicolas Dessaigne 1:51 Uh, w- y- where did that come from?

Matt 1:54 That's a good question. I mean, uh, we... there are actually four co-founders, um, and Dan and James are, like, really good at UX and design. Um-

Eric 2:01 It might be the Britishness, maybe. I don't know.

Matt 2:03 [laughs]

Nicolas Dessaigne 2:03 [laughs]

Matt 2:04 Could be. Could be.

Nicolas Dessaigne 2:04 The Jony Ive kind of thing?

Eric 2:05 Yeah.

Matt 2:06 [laughs]

Nicolas Dessaigne 2:07 [laughs]

Eric 2:08 Jony Ive.

Matt 2:08 Yeah, I guess we really value design. We think the best developer tools actually care about design. And when we say design, I'm not talking just visual design. I also think that, like, developer experience is actually about design, um, like, designing the experience so that it's easy to succeed.

Nicolas Dessaigne 2:22 Hm.

Matt 2:23 Um, and we spend... Probably the hardest conversations we have, um, around the product are, like, how to design this, like, specific SDK function and make it so dev- it's very, very hard for developers to, like, fail when they use it.

Nicolas Dessaigne 2:36 Hm.

Matt 2:37 Um, so we're very, very careful about that. Um-

Nicolas Dessaigne 2:40 I remember during the batch how much time you spent specifically on just, just the landing page and trying to, trying to communicate what Trigger did. You had this, uh, you had this, like, code snippet on the landing page.

Matt 2:56 [laughs]

Nicolas Dessaigne 2:56 It was the first thing you saw when, when you open, when you opened it up, and you spent a long time trying to, I, I think, highlight specific sections, and, um, that was, that was the hardest I had seen anybody think about the first five seconds that a developer has, uh, learning about your tool.

Eric 3:12 And we maybe spent too much time on it. [laughs]

Matt 3:15 [laughs]

Eric 3:15 But yeah, the, the code snippet thing, it's, it's still there. Like, we, we sort of... I always hate going to someone's, like, a developer website, and, like, you're having to scroll, scroll, scroll, and there's no code. And I feel like I just wanna see the code. Like-

Nicolas Dessaigne 3:26 Yes

Eric 3:26 ... like, tell me what the code looks like.

Nicolas Dessaigne 3:27 Yes.

Eric 3:27 'Cause that's sort of, like, what I... That's how I get into the product, right? How I see, like, oh, is this something I wanna use?

Nicolas Dessaigne 3:33 Yes.

Eric 3:33 It's like, what does the code look like? Is it gonna be very complicated? Is it gonna be very simple? And I'd actually think that's sort of, like, code first approach. Like, that was one of our big things, even though you're, you know, you're talking about design. But, like, it really, it is, like, we thought of the code that we wanted users to write sort of first, and then we sort of worried about, like, how do we actually get it all to work.

Nicolas Dessaigne 3:52 So you started with a, with a sort of Zapier for developers. Um, and, uh, you had, as I recall, like, a, a great time in the batch. You raised money coming out of the batch.

Matt 4:02 Mm-hmm.

Nicolas Dessaigne 4:03 Uh, the idea has obviously changed a lot since then. Uh, maybe talk a bit about how it evolved over time. Like, what did you learn that informed the direction you ended up taking things?

Matt 4:14 Yeah. So I think the first use cases, because we were kind of positioned as this, like, Zapier for developer things, were kind of what I would call back office, um, things. So things like internal teams were building tools using us. Um, they were, like, replacing what Zapier is good for.

Nicolas Dessaigne 4:28 Hm.

Matt 4:29 Um, but we pretty quickly realized that the best use cases were where people were actually, like, baking us into their products.

Nicolas Dessaigne 4:36 Hm.

Matt 4:37 Um, and so version two of, uh, Trigger, which we released, like, a couple of month- a few months after the batch, um, was more focused on that. So it was more about adding, uh, async tasks to your existing product. So less back office use case, more embedded in your product.

Nicolas Dessaigne 4:53 So what's an example of, of kind of the, the back office background task you had in mind when you designed the first version, and, and, and what's an example of a baking this straight into the product-

Matt 5:05 Yeah

Nicolas Dessaigne 5:05 ... use case?

Matt 5:06 So back office stuff would be, like, you're automating stuff with... For, for developers for, like, automating GitHub, um, maybe doing-

Nicolas Dessaigne 5:13 Ah

Matt 5:13 ... some marketing use cases. So it's like the, it's the-

Eric 5:18 Sales.

Matt 5:19 Yeah. It's like biz ops and sales and marketing, the stuff that, like, a lot of people use Zapier for, um, and, like, NAN and things like that. And then baking into your product, it's more that you, um, you are actually executing code that's part of your product. So rather-

Eric 5:33 Like doing something for your users. You know, your user has done something, and now you wanna do something sort of in the background asynchronously, like process a document or, um-

Nicolas Dessaigne 5:41 Oh, interesting

Eric 5:42 ... you know, encode a video-

Nicolas Dessaigne 5:42 Interesting

Eric 5:42 ... or do something like that. It's like, so it's, it's in the hot path of your, like, app that provides value directly to your users-

Nicolas Dessaigne 5:50 Sure

Eric 5:50 ... not providing value to your team.

Nicolas Dessaigne 5:53 Fair enough.

Eric 5:53 Yeah.

Nicolas Dessaigne 5:54 You, you've basically given developers a really simple way to spin up a background process and-

Eric 5:58 Yeah

Nicolas Dessaigne 5:58 ... get something done.

Eric 5:59 Yeah, and run it through queues and make sure, you know, things are reliable and retry and idempotent and all that sort of thing. Yeah.

Nicolas Dessaigne 6:06 That makes, that makes a lot of sense.

Matt 6:08 And I think because there, there was this shift to serverless, um, over the previous, like, 10 years, um, that stuff became a lot harderUm, like it's really good at doing short-lived like request response, um, stuff, but not very good for long-running tasks. And so it kind of left like a big gap. Um, and that's what we were filling basically.

Nicolas Dessaigne 6:31 How well was this adopted, right?

Matt 6:33 [laughs]

Nicolas Dessaigne 6:33 Like were you, were you starting to grow at this-

Eric 6:34 Yeah

Nicolas Dessaigne 6:34 ... at this point?

Matt 6:36 It w- it did okay.

Eric 6:36 It, it was okay.

Matt 6:37 It did okay.

Nicolas Dessaigne 6:38 Okay.

Matt 6:38 But it definitely wasn't like product market fit. Um-

Nicolas Dessaigne 6:40 Fair enough

Matt 6:41 ... and so, and I think partly it's because it just wasn't solving the problem that well.

Nicolas Dessaigne 6:46 Huh.

Matt 6:46 Like we f- there was definitely like demand-

Nicolas Dessaigne 6:49 Mm

Matt 6:49 ... but we, so I think the market was there, but the product didn't really match. And I think two really important things changed. Uh, one was that, uh, AI started to take off. Um, and it just so happens that async tasks are, background async tasks are very, very useful for AI. And then the other thing is that we realized that our solution wasn't very good, and it wasn't very good because it involved you writing still quite messy code.

Nicolas Dessaigne 7:16 Mm.

Matt 7:16 Um, and we decided to actually start executing code, um, and that was version three of the product, which we released in, uh, June 2024.

Eric 7:27 Yeah, summer '24.

Matt 7:28 Um, and that's when stuff started to really take off.

Nicolas Dessaigne 7:30 So it sounds like version two, you gave developers an SDK, but they were still executing these background jobs on their infrastructure.

Matt 7:41 Yeah.

Nicolas Dessaigne 7:41 And in version three you offloaded the e- the execution entirely to your infrastructure.

Matt 7:47 Yeah.

Eric 7:47 Yeah. We actually did a poll when we were doing version two, and I asked, you know, our customers like, "Where is the code executing it? Are we executing it or are you executing it?" I think 60% thought we were executing it, [laughs] so like they already thought we were doing it.

Nicolas Dessaigne 8:01 [laughs]

Eric 8:02 Yeah. So we were like, "Maybe we should do it." [laughs]

Matt 8:06 [laughs]

Nicolas Dessaigne 8:07 Uh, and so, and so I was y- I was going to ask if you had trouble kind of convincing developers to trust you to handle this execution, but what-

Matt 8:13 Yeah

Nicolas Dessaigne 8:13 ... it sounds like is they already thought you were doing it. [laughs]

Eric 8:15 There, there was definitely like a sub-

Matt 8:16 We thought we would

Eric 8:17 ... there was a subset of people-

Matt 8:18 Yes

Eric 8:18 ... who were, who did like that model and understood it.

Matt 8:20 Yes.

Nicolas Dessaigne 8:20 Mm.

Eric 8:21 And, and then when we switched, they, we were like, "Yeah, there's, you go pick these other people who sort of do that use case for you."

Matt 8:28 Understood.

Eric 8:28 But almost everyone else, like w- that's what they wanted. They, they, they wanted it to actually run for a very long time and, and have us execute it and manage it, and not do it themselves.

Nicolas Dessaigne 8:39 Makes sense. So, so version one, just recounting here, version one was Zapier for developers. Version two was an SDK for, for doing what felt like an asynchronous background job.

Matt 8:49 Yeah.

Nicolas Dessaigne 8:49 Version three was, was just an SDK for doing actually an asynchronous background job.

Eric 8:53 Yeah, it was the SDK, the platform, and the infrastructure.

Nicolas Dessaigne 8:56 And, and so you launched that. Is, do we have product market fit yet?

Matt 8:59 Uh, yeah.

Eric 9:00 Yeah.

Matt 9:00 Pretty much straight away.

Nicolas Dessaigne 9:01 Okay.

Matt 9:02 Um, like we were growing super strongly month over month, like revenue was just going up, and like that's been the story ever since basically.

Nicolas Dessaigne 9:09 Okay, so you've, you launch version three and you, you just immediately see developers starting to pick this up and using it. Are you making money at this point?

Matt 9:16 We s- we, initially it was like a beta and we weren't charging, but then like a couple of months later we started charging for it. Um, and yeah, we, the revenue growth was like plus 30% every s- at least every single month.

Nicolas Dessaigne 9:29 Wow.

Matt 9:29 For like a, for a l- a long time it was that much.

Nicolas Dessaigne 9:31 Wow. Wow. Um, you're an open source product.

Eric 9:35 Mm.

Matt 9:36 Mm.

Nicolas Dessaigne 9:36 Has that, has there, has it been a challenge, like w- was figuring out the business model and, and how to actually make money a challenge here?

Matt 9:44 Not really because we're, because we're running compute-

Nicolas Dessaigne 9:48 Mm

Matt 9:48 ... and most developers don't wanna actually think about infrastructure that much. We just have a cloud offering.

Nicolas Dessaigne 9:54 I guess your whole product, the pitch is you don't have to think about infrastructure.

Matt 9:57 Yeah.

Eric 9:58 Yeah, yeah.

Matt 9:58 I think there are a lot of interesting questions, um, around what is included in the open source project and what isn't. Um, and I think this is a debate that lots of open source founders have.

Nicolas Dessaigne 10:10 Yes, I hear it a lot. Where did you land on that?

Matt 10:13 At the moment, so we're Apache 2 open source, so we're like very permissive.

Nicolas Dessaigne 10:17 Mm-hmm.

Matt 10:17 And pretty much everything is in the open source product at the moment. Um-

Eric 10:21 Yeah, the only thing that's not in the open source is sort of, well, how do you actually like manage all the servers and spin them up and like re- like, like actually get the, all the pieces in place.

Matt 10:31 Yeah.

Eric 10:31 That's like the whole Kubernetes layer, you know.

Matt 10:33 Understood, yes.

Eric 10:34 So that, that sort of stuff isn't, isn't quite, uh, open. Although we have a Kubernetes helm chart and people can run it, but it's, it's hard to scale it really well-

Matt 10:43 Makes sense

Eric 10:43 ... basically.

Nicolas Dessaigne 10:43 Interesting. So, so, ma- and maybe can you give us an example here just again to make this concrete, like what was-

Matt 10:49 Yeah

Nicolas Dessaigne 10:49 ... what were some of these early use cases, uh, that, that your, your customers were, were sort of pulling your product towards?

Matt 10:56 Yeah. So, so there's a company called icon.com, and they're basically replacing video ad companies.

Nicolas Dessaigne 11:01 Mm.

Matt 11:02 So you go on and you upload some assets of your product, and then you describe what you want your video ads to be like.

Nicolas Dessaigne 11:08 Mm.

Matt 11:09 And then they use Trigger to like process the assets, so classify them, and then also just like generate videos.

Nicolas Dessaigne 11:16 Mm.

Matt 11:16 Um, and you get like hundreds of adverts spat out, and then you can give feedback on those adverts, and eventually you post them to TikTok and Instagram. Um, so that's like a good example of how to use Trigger.

Nicolas Dessaigne 11:27 And, and in that, in that f- workflow, the end user uploads all of these assets, and then Icon hands the job to Trigger.

Matt 11:39 Mm-hmm.

Nicolas Dessaigne 11:39 Trigger does this sort of long-running workflow of processing all of the things they've uploaded, then generating new assets.

Matt 11:48 Mm-hmm.

Nicolas Dessaigne 11:48 Uh, and then, you know, you send a like, "Hey, we're done," and it, you hand it back to, to, to Icon?

Matt 11:53 Yeah. In, in real time you can see all the assets being processed, but then the, that's kind of, there's kind of two p- in fact, we talk about this a lot.

Eric 12:00 Yeah.

Matt 12:00 There's kind of two parts to building successful agents. There's the like context that you need.

Nicolas Dessaigne 12:05 Yes.

Matt 12:06 And then there's the actual, like the moment where you're doing something with that context. And so this is a perfect example, because in the case of Icon, the context is like all of the assets.

Nicolas Dessaigne 12:15 Mm-hmm.

Matt 12:15 Um, and then the actual generation phase is where the user describes what kind of advert they want to make.Um, and it pulls the correct context, and then it generates new data, like maybe it generates AI actors, um, for their video. Um, and it, in r- in both cases, you want real-time feedback.

Nicolas Dessaigne 12:36 Uh, so it, it sounds like Trigger's actually being used to kind of do the agent loop here, where the users, once, once the context has been uploaded, the user's able to say, "No, I want, you know, a video ad that does X," or, "I, you know, I'd like one that looks like this."

Matt 12:48 Yeah, and human in the loop, by the way, is like a critical part of that.

Nicolas Dessaigne 12:51 Interesting.

Matt 12:51 Um, so you have this thing running, and then it can pause and ask for feedback.

Nicolas Dessaigne 12:57 Yes.

Matt 12:57 Um, it could be from a human, it could be from another agent.

Nicolas Dessaigne 12:59 Yes.

Matt 12:59 Um, and like we will pause the compute, um, until the feedback is received, and then we'll continue it.

Nicolas Dessaigne 13:05 It almost seems like this thing that you had been building, um, for, for almost two years at that point, was just perfectly designed for a use case that didn't actually exist at the time.

Matt 13:17 [laughs]

Nicolas Dessaigne 13:17 Uh, like these long-running agent loops, where you need to frequently pause, check back in, collect more context from the user. Uh, none of that was around when you started this, but you're n- you, you find yourselves, like, in the perfect position to, to solve this problem for developers.

Matt 13:31 Yeah, we definitely got lucky. [laughs]

Nicolas Dessaigne 13:33 [laughs]

Eric 13:33 I think that's what he was saying. [laughs]

Matt 13:36 [laughs]

Nicolas Dessaigne 13:36 Let's fast-forward to today. Um, how much of your usage is running agent workflows like the one that you just described?

Matt 13:44 Yeah, it's like over 90%.

Nicolas Dessaigne 13:46 Wow.

Matt 13:46 Yeah.

Nicolas Dessaigne 13:47 Wow. Um, what are some other interesting ways that you've seen developers using this?

Matt 13:53 Yeah, so another customer is Magic School.

Nicolas Dessaigne 13:56 Okay.

Matt 13:56 And they basically provide tools for teachers, um, and students. Uh, so they're a really successful ed tech company.

Nicolas Dessaigne 14:02 Okay.

Matt 14:03 And so they, their tools basically make it so that teachers can run their lessons more efficiently.

Nicolas Dessaigne 14:08 Mm-hmm.

Matt 14:08 So it helps with lesson planning-

Nicolas Dessaigne 14:10 Mm

Matt 14:10 ... asking questions, marking homework. Um, and on the student side, it helps them make sure that they don't fall behind.

Nicolas Dessaigne 14:16 Fair enough. And they're using, these are, these are all done through, through kind of agent workflows, and so it's the teacher chatting, uh-

Matt 14:23 Yeah

Nicolas Dessaigne 14:23 ... with an agent-

Matt 14:24 Still running in the background

Nicolas Dessaigne 14:25 ... to do lesson planning. And, and is, is Trigger handling the entire agent, uh, execution?

Matt 14:31 Yeah. The actual agents execute using Trigger, yeah.

Nicolas Dessaigne 14:33 All in, in Trigger. Okay, got it.

Eric 14:35 Yeah.

Nicolas Dessaigne 14:35 And what else?

Eric 14:35 And another one is, like, uh, Scrapybara. Um, they were using us to, they build, like, a code- coding agent, basically. So think, like, uh, you go into their website and log in, and you hook up your GitHub, and then you can-

Nicolas Dessaigne 14:46 Yes

Eric 14:46 ... like, have it build your features and stuff. And so all that, like, pulling down the code and executing it and, like, evaluating it and, like, running LLMs to change it and submit the commits to GitHub and all that, that was all running on Trigger.dev, yeah.

Nicolas Dessaigne 14:58 Are you giving developers access to the shell itself? Um, or, or, or does a developer interact with all of this through a, uh, like a, a l- a language-specific SDK?

Eric 15:11 Yeah, so they, they have full, the full power over the machine, basically. So they can do shell stuff. They can do Python. I mean, we are, [laughs] we are a TypeScript-first, like, framework.

Nicolas Dessaigne 15:20 Right.

Eric 15:20 But we, we, what, what is sort of output is, like, a whole machine image that you've customized.

Nicolas Dessaigne 15:26 Got it.

Eric 15:26 So you can load anything on there that you want, and then you can spin it up and, and run FFmpeg or Puppeteer with Chrome, and, which a lot of people do. A lot of people use us to, like, run, uh, you know, Chromium in the background and, um, do, like, computer use stuff.

Nicolas Dessaigne 15:41 Got it.

Eric 15:41 Um, and so yeah, we gi- we give you, like, the full power of the, sort of the machine. Um, but you interface it through with, like, a nice, like, TypeScript, uh, framework, 'cause that's sort of the language of, like, AI applications-

Nicolas Dessaigne 15:53 Mm-hmm

Eric 15:53 ... or is TypeScript. Um, I know the Python people might not like that, but [laughs]

Nicolas Dessaigne 15:57 [laughs]

Eric 15:58 Um, but yeah, so you, you, you can do that. And we're, what we're doing is sort of, um, you know, the pausing and resuming kind of primitive, we're, we are bringing that out and, like, giving people the power to, like, programmatically do that. Um, and we think that's, like, really where the future of, like, computing is, is basically pr- programmatically being able to do this type of snapshot and restore mechanism. 'Cause what you really wanna do is, is kind of freeze the compute where it was at, and you, the whole state of the compute, the CPU, the memory, the file system, and then you wanna say, when whenever something else happens, like, please re- resume, and now everything that was there before is still there, and you can kind of pick up where you left off. Traditionally, you would have to rehydrate all the state, you know. Um, and that's, like, a tricky thing to do, and t- like, tons of different time, ways of doing it. Um, so we think the whole, uh, building everything off of checkpoint restore is, like, the, the way of the future, basically.

Nicolas Dessaigne 16:52 It, it sounds like it's like an operating system scheduler, but-

Eric 16:55 Yeah

Nicolas Dessaigne 16:55 ... operating over different containers in different cloud instances that I never have to think about as a developer.

Eric 17:01 Right. And it, it is, yeah, it's, it's [laughs] it's tricky to, to, to do. Um, but we're trying to basically, like, democratize that technology, 'cause it is quite, like, low level, and there's a lot of parts to getting it right, basically.

Nicolas Dessaigne 17:14 One of the other big things that's changed in the last year is just there are a lot of people now building software that don't have a traditional software background. Um, and I'm curious whether you see a split in your customer base, uh, between kind of traditional devs, uh, using this the way that they would any other infrastructure provider, versus, like, vibe coders who, who, who have just picked it up.

Eric 17:37 Yeah, I mean, I think the vibe coding thing, we, we felt a real big, like, bifurcation in the vibe coder versus traditional coders, like, six months ago and a year. But I feel like that, that has, like, collapsed almost.

Nicolas Dessaigne 17:50 S- say more. W- what have you seen?

Eric 17:52 Well, because, well, because they're, they're using Opus 4.5 now.And it's better. And so there's, they are able to like use us well.

Nicolas Dessaigne 18:00 Ah.

Eric 18:00 Right?

Nicolas Dessaigne 18:00 Okay.

Eric 18:00 Whereas before we were seeing a lot of people using us very poorly.

Nicolas Dessaigne 18:03 Mm.

Eric 18:03 And a l- we have done some, a lot of work as well in the last six months, so maybe part of it is that to basically be more like LLM friendly.

Nicolas Dessaigne 18:11 Yeah.

Eric 18:11 We, you know, built a MCP server. All, all of our docs, like, we've like kind of fleshed them out better for agents and-

Nicolas Dessaigne 18:18 Skills

Eric 18:19 ... and skills and like we've done all those things. So that's probably part of it. But I think just because the, also the agents, the coding agents are getting much better-

Nicolas Dessaigne 18:25 Mm

Eric 18:25 ... that they're able to use us better. And so they're, our users, the vibe coders are running into less issues.

Nicolas Dessaigne 18:31 Got it.

Eric 18:31 So they're actually doing better now.

Nicolas Dessaigne 18:32 So when you say the bifurcation has, has gone away-

Eric 18:34 Yeah

Nicolas Dessaigne 18:34 ... you don't mean that they've all dropped off and stuff?

Eric 18:36 No, no, they're still there.

Nicolas Dessaigne 18:37 What you mean is like they're, you're, you're less able to distinguish between-

Eric 18:40 Yeah, yeah

Nicolas Dessaigne 18:40 ... the, like, serious devs versus-

Eric 18:42 Yeah

Nicolas Dessaigne 18:42 ... the vibe coders.

Eric 18:42 Yeah, because we could tell before.

Nicolas Dessaigne 18:44 Hm.

Eric 18:44 It was very obvious.

Nicolas Dessaigne 18:45 Hm.

Matt 18:45 Yeah, the support. There were like two types of support-

Nicolas Dessaigne 18:48 Yes

Matt 18:48 ... that we would [laughs] have. There'd be people who clearly had never written software before. Um, and especially when this first started happening, they were struggling because the product was definitely not built for them.

Nicolas Dessaigne 18:58 Yes.

Matt 18:59 Um, and we did a lot of work to make that better. And then there were questions from, like, much more advanced users. But now it's, like, all blurred.

Nicolas Dessaigne 19:07 Wow.

Matt 19:08 Yeah.

Nicolas Dessaigne 19:08 As a developer myself, um, one of the, like, great luxuries of this moment is that every dumb question I have, I can just ask Claude Code for an answer first, uh, right? And that, that, you know, reaching out to a, a human line of support feels like a last resort for me. Have you noticed, uh, just a change overall in how much people are using or how little pe- people are using your customer support?

Matt 19:31 It's a little bit hard to untangle because we've been growing fast.

Nicolas Dessaigne 19:34 Yeah.

Matt 19:34 And so, like, the support has been going up. I think it's been going up a lot less fast than it was.

Nicolas Dessaigne 19:39 Hm.

Matt 19:40 Um, like, I, I guess if you, if you were to plot our, um, like, number of customers and executions and, uh, revenue and support, I think there's, there's, the lines are diverging. Um-

Eric 19:51 But, but one big benefit that we have is because we're open source, a lot of times our customers will just, like, pull down the repo and ask it questions about, like, how to use us based on just the repo.

Matt 20:02 Yes.

Eric 20:02 Our code is out, out there, so they can do that. But then also sometimes they'll contact us and be like, "This is the bug. Like, we found it." Like Claude Code found it, and here's the write-up, and like I can go in Slack and just say, "Claude, like, fix this," and then Claude fixes it in the background and then pushes the PR to a repo.

Nicolas Dessaigne 20:19 [laughs]

Eric 20:19 And it's like, "Okay, it's fixed now." But so, like, the quality of support, it, for the people who are using these tools, like, well-

Nicolas Dessaigne 20:25 Yes

Eric 20:25 ... has definitely, like, increased.

Nicolas Dessaigne 20:27 That's really fascinating.

Eric 20:28 Yeah.

Nicolas Dessaigne 20:28 Uh, and actually it didn't occur to me, uh, how much of an advantage open source companies have, uh, because all of their code, which is the source of truth about how the product works, is-

Eric 20:38 Yeah

Nicolas Dessaigne 20:38 ... available to-

Eric 20:39 Yeah, because they can read the test, right? [laughs]

Nicolas Dessaigne 20:41 Yes. Yes, yes, yes.

Eric 20:41 Which is usually what you do is like read the test and figure out what, what actually is going on. Um...

Matt 20:45 We, we talk about that a lot as the, there's kind of two users now. There's the human user, uh, who wants to build something, but also, like, the LLM is a user of Trigger. And so, like, how do we get LLMs to love using Trigger and, like, choose it? Um, and I think being open source is a massive advantage there because we have a much bigger footprint on the internet-

Nicolas Dessaigne 21:08 Yes, yes

Matt 21:08 ... than if we were closed source.

Nicolas Dessaigne 21:10 So, so maybe talk a bit more about that. Um, how much, how much time and attention do you spend in your marketing, uh, just focused specifically on agents as users?

Matt 21:22 Probably not as much as we should. [laughs]

Eric 21:24 Yeah. How do you market to an agent? That's...

Matt 21:26 [laughs]

Nicolas Dessaigne 21:27 How often do you think your customers, uh, sort of are consciously aware that they're choosing Trigger.dev when they first become users, versus how often is it, you know, I want to do X and Claude Code, uh, you know, reasons and, and decides that Trigger.dev is the best way to do that?

Matt 21:46 Yeah. I think at the moment, the stage we're at right now is that, um, AI is recommending people to use Trigger, but the human still has to, um, like, actually, like, go through the sign-up flow.

Nicolas Dessaigne 21:57 Fill in the trigger.

Matt 21:58 Yeah. [laughs]

Eric 21:59 [laughs]

Matt 22:00 But, um, we've been talking a lot recently about how we want it to be like an end-to-end experience. Um, so Claude can just, like, decide to use Trigger, and it will just, like, spin up an account and start actually, like, using us.

Nicolas Dessaigne 22:14 Oh, wow.

Eric 22:14 It can almost do it. It's just the account part is probably-

Nicolas Dessaigne 22:17 Yeah

Eric 22:17 ... missing. But everything else it can do.

Nicolas Dessaigne 22:19 Wow.

Eric 22:19 Yeah.

Nicolas Dessaigne 22:19 That's really interesting.

Eric 22:20 Yeah.

Nicolas Dessaigne 22:21 So the onboarding is... You're, you have to think about the agent onboarding experience maybe before the human onboarding experience.

Eric 22:26 Yeah, I think, but easily within the next few years, everyone will have to have some sort of agent.

Nicolas Dessaigne 22:32 Do you guys use your own product to build internal tools?

Matt 22:38 Yeah, we do. Um, and we're building, like, we're building, like, a support agent at the moment.

Nicolas Dessaigne 22:44 Yes.

Matt 22:44 Um, so we, we use, we use some existing support tools. Um, and I think we could do like a better job when we do it very specifically for our use case.

Eric 22:55 Although one of the cruel ironies-

Nicolas Dessaigne 22:57 Yes

Eric 22:57 ... of building a thing like this is that you keep wanting to use it, [laughs] but then every time you use it, you go, "Oh, that, we have to make that better." So you end up stopping using it, and you build, like, to make it better. Because it's really hard to, like, force yourself to, like, not see an issue and go and fix it.

Nicolas Dessaigne 23:13 I mean, that's

Nicolas Dessaigne 23:14 the best kind of feedback loop is-

Eric 23:15 Yeah

Nicolas Dessaigne 23:15 ... uh, to be totally honest. And I, and I guess, like, there's this really virtuous cycle where, uh, the more you use the product, uh, the better pr- the product becomes, and the easier it is to automate a bunch of work within your own company.

Eric 23:27 Yeah.

Nicolas Dessaigne 23:27 Um, maybe let's talk a little bit about that. How have you thought about hiring in this world?

Matt 23:32 Yeah. So, so we, we, uh, raised our Series A, like, last, like, November time. Um, and you know, our hiring plan has materially changed since then. Um, we're definitely not going to be hiring as many engineers as we were going to. As quickly-

Nicolas Dessaigne 23:50 Oh, meaning in November you had planned to hire a lot of engineers, and at this, at this moment, which is, is February-

Matt 23:56 Yeah

Nicolas Dessaigne 23:56 ... early February 2026-

Eric 23:57 Post Opus 4 point post.

Matt 23:58 Yeah.

Nicolas Dessaigne 23:58 Yes.

Matt 23:59 Yeah. Basically, with the release of-Uh, better, like the better planning tools and Opus 4.5

Eric 24:05 And Claude code improvements

Matt 24:06 Yeah

Eric 24:07 Yeah

Matt 24:07 Like our productivity per engineer is, I don't know, 5X-

Eric 24:10 Mm

Matt 24:11 ... 10X what it was before. Um, and so we have hired people, um, and we will hire some more people, but definitely much less aggressively than we were going to.

Nicolas Dessaigne 24:22 What have you done in order to make it easier for the team to work with Cl- like, you've been working on this product for two and a half years, uh, you know, and this is not a, this is no longer a greenfield project.

Matt 24:34 Mm.

Nicolas Dessaigne 24:34 I'm assuming you have like a, a reasonably large code base now that is fairly hairy. How, what have you done to make it easier for the engineers on your team and yourselves to work with coding agents in that code base?

Eric 24:46 Um, to be honest, you don't have to do that much anymore. [laughs]

Nicolas Dessaigne 24:51 [laughs]

Eric 24:51 Like it's sort of the, you know, the, they're so good.

Nicolas Dessaigne 24:54 Yeah.

Eric 24:54 The agents are so good, and we, the, you, i- it will, it will figure it out by itself. I mean, obviously we do the obvious stuff, right? Like we build, you know, we have all the Claude MD and Agents MD files and all, all that sort of thing, and we've, we've, um, developed our own like internal sort of tools and skills-

Nicolas Dessaigne 25:11 Mm

Eric 25:11 ... and things like that that give people like re- really, you know, stuff that we commonly do, and stuff that's like very specific about our repo and all that. But it's just, yeah, y- you kind of like, you- we almost were like lazy to the point of like letting the [laughs] models like catch up, and now, you know, we don't have, there's not a bunch of stuff we have to do to get [laughs] to be good. It's just already good.

Nicolas Dessaigne 25:31 Fair enough.

Eric 25:31 Yeah, yeah.

Nicolas Dessaigne 25:32 Fair enough. So y- faced with the decision of whether to do a bunch of work to make it better or just wait for the models to get better, you're just like, "Ah, let's just, uh..."

Eric 25:39 Yeah, yeah.

Nicolas Dessaigne 25:39 [laughs]

Eric 25:39 Got it. I mean, w- it wasn't on purpose, but yeah.

Nicolas Dessaigne 25:42 [laughs]

Eric 25:42 Let's put it that way. [laughs]

Nicolas Dessaigne 25:43 Um, okay. So, so you, uh, you are still hiring engineers.

Matt 25:47 Yes.

Eric 25:47 Yes.

Nicolas Dessaigne 25:47 Not as aggressively as you had planned to a couple months ago. What about the other functions in the company?

Matt 25:53 Yeah. So, um, marketing obviously. Um, so we've recently hired a dev rel, um, and we're looking for a content engineer.

Nicolas Dessaigne 26:02 Mm-hmm.

Matt 26:03 Um, so we're, we're about to start publishing a lot of content around how to build great agents.

Nicolas Dessaigne 26:10 Got it.

Matt 26:10 Um, and just, it will be just general, like it won't be specific to Trigger. It's just we think there is a lot of value we can add based on like looking at our customers and our experience.

Eric 26:21 Yeah, and part of the thing that like we're evaluating even programmers on or engineers on, right, is, well, how well do they use these tools? Like how well are they... Like y- you're, we're not like looking for people who, um, you know, necessarily can go and like write something from scratch without using the tools. Like we actually know, we wanna know, like how are you using them? How are you leveraging them? Like that's the skill now, that's the actual job, is like how are you properly leveraging the agents to like do y- do your whatever you're work.

Nicolas Dessaigne 26:50 How, how do you test that? Do you, do you, is this baked into your interview process?

Matt 26:54 Yeah, so the, like one of the later stages in the interview is, um, we do a trial day. So we pay you to work for us for a day.

Nicolas Dessaigne 27:02 Hmm.

Matt 27:02 And, um, you are encouraged to use the normal tools you would use, and if someone isn't using AI tools well, then like we're not gonna hire them.

Nicolas Dessaigne 27:11 And are you expecting them to ship something to production on their, on their, their day?

Matt 27:15 No, we don't. Um-

Eric 27:17 If they want to.

Nicolas Dessaigne 27:18 I mean, maybe you should. Yeah.

Eric 27:19 [laughs] Yeah.

Nicolas Dessaigne 27:19 The interview is just a job. That's k- that's kind of interesting.

Matt 27:22 I think it's like, um, it's the absolute opposite of LeetCode interviews. Like LeetCode interviews are like these things that are, you kind of, most people, if we're honest, revise for them, and then you do them obviously without AI, and it's like, how practical can you be? Like how can you ship really high quality software? And if you're not using AI to do that, I think you're

Matt 27:41 crazy.

Nicolas Dessaigne 27:41 A lot of the criticism around, uh, uh, the use of AI coding agents in, in large code bases is that they t- they just tend to produce slop. Like unwieldy, hard to maintain code and, and maybe unwieldy product experiences too. And you guys strike me as, as founders who more than most care about th- the really small details of how your product looks and feels. How do you and your, your engineers use coding agents to produce such high quality software? Have you had to think about that?

Eric 28:10 Mm.

Matt 28:10 Yeah, definitely. I think there's a lot of aspects to that. One is, um, I think you still need like good components. Um, if we're talking about like the front end side of things-

Nicolas Dessaigne 28:19 Okay

Matt 28:19 ... I still think you need like high quality components. You don't want AI to create like the component from scratch each time.

Nicolas Dessaigne 28:24 Hmm.

Matt 28:25 Um, so like we have a really strong design system, um, that we use, and we obviously adapt it over time. Um, I think our bottleneck right now is actually more on the review side. Like we can generate a ton of code-

Nicolas Dessaigne 28:37 Yes

Matt 28:37 ... um, but the quality control and review, and that review is both like reviewing whether the code is good and then like kind of de-slopping it, um, and using AI to do that, but also, um, like UX.

Nicolas Dessaigne 28:52 Yeah. And it, so do you have a process for this? Is it, is it just you, you know, the founders kind of going through everything with a fine-tooth comb? Have you built a pi- a pipeline to try and, and make that easier?

Matt 29:04 We use, we've been like evolving this a lot. We use good code review tools.

Eric 29:08 Yeah.

Matt 29:08 Um-

Eric 29:08 With Dev- with the new Devon, uh, code review is very good for this.

Nicolas Dessaigne 29:11 Okay.

Eric 29:12 Um, and we kind of combine a few different ones. Uh, it depends on what you're doing, like which kind of, um, like problems you would run into with the, the, like using the code agents. I think one of the mistakes you might make by like using coding agents is just viewing them as something that would, uh, write your code.

Nicolas Dessaigne 29:30 Mm-hmm.

Eric 29:31 But what you can use them for is like a bunch of stuff around writing your code.

Nicolas Dessaigne 29:35 Yeah.

Eric 29:35 So like maybe, uh, oh, you might think of like, oh, I wanna like start using this other thing. We'll benchmark it against this per- the thing we're using now, right? Uh, see how it uses the CPU. Like is it more efficient, less efficient?

Nicolas Dessaigne 29:49 Mm.

Eric 29:49 Like that's a lot of work-

Nicolas Dessaigne 29:50 Hmm

Eric 29:50 ... usually.

Nicolas Dessaigne 29:51 Yes.

Eric 29:51 But like while you can do that with AI within a few minutes.

Nicolas Dessaigne 29:54 Yes.

Eric 29:54 And now you've, you've now increased your knowledge about like what is a good thing to include in your code or not.And you would probably wouldn't have done that before. You might have just like YOLO'd it and just been like, "I, well, hopefully it'll work better," right? And we have like our prod sort of observability stuff, and it'll tell us, right? But now it's like, well, I can move a lot of that analysis earlier because it won't take me a week to write the benchmark.

Nicolas Dessaigne 30:16 Yes.

Eric 30:17 Um, so it's a lot of things like that where it's like imagine if you had like an army of people, and you could just be like, "Oh, can you check that for me? Do that for me. Do that."

Nicolas Dessaigne 30:23 Yes.

Eric 30:23 Go run... Like, see what's happening in production with this. Like, is it safe if I change it to that? Like, you can actually do all that stuff now, and that's sort of what we're like looking for, like people using agents not just to like produce the code.

Nicolas Dessaigne 30:34 Yes.

Eric 30:35 It's like do the job of like a, a, a staff engineer-

Nicolas Dessaigne 30:38 Yes

Eric 30:38 ... but with like a ton of other engineers that are helping them.

Matt 30:42 I think on the back end, there's also just like tests. So if we're honest, like a year ago, we probably didn't have that many tests in the code base. We now have a lot of tests, and whenever we're building back-end features, like creating like verifiable... Like the agent needs to know if it's succeeded.

Nicolas Dessaigne 31:00 Yes.

Matt 31:01 Um, and so it, we, we have a, it write a lot of tests. Um-

Eric 31:05 For the front end, we have James. [laughs]

Nicolas Dessaigne 31:09 [laughs]

Matt 31:09 Yeah, I mean, that-

Nicolas Dessaigne 31:09 'Cause the tests

Matt 31:09 ... that seems to be the thing that everybody is struggling with, and, and I suppose James is the, is the ultimate ardi- arbiter is, uh-

Eric 31:16 Unfortunately, yeah,

Eric 31:17 we can't share him.

Matt 31:17 Fair enough.

Eric 31:18 [laughs]

Matt 31:19 Well, there's your moat. [laughs]

Eric 31:21 [laughs]

Nicolas Dessaigne 31:21 Right now at YC, we're gearing up for the spring batch, and that means we're gonna have a whole bunch of new founders coming in, uh, that are in a similar position to the one that you were in at the beginning of winter '23, uh, three years ago. Do you have any advice?

Eric 31:34 I mean, it's gonna sound a lot like the YC advice, but, uh, I really think shipping early is, is probably the most important thing you can do.

Nicolas Dessaigne 31:41 Mm.

Eric 31:41 And, and actually... 'Cause once you ship, then you can actually learn what people want.

Nicolas Dessaigne 31:46 Yeah.

Eric 31:46 It's easy like to sort of, yeah, talk to people before you're shipping, and obviously that's very important, but shipping and having something out there and having something that people can pay for and having feedback and bugs and all that stuff, like you- you'll learn sort of the constraints of whatever you're trying to do, and maybe you'll learn that like, oh, you had some downtime, and no one emailed you about it, or no one sent you a message about it. So it's like, oh, maybe we're not building something important.

Nicolas Dessaigne 32:08 Mm.

Eric 32:08 It's like stuff that you just, you, you can only learn by shipping. So it's not like anything super like innovative, but it is just like please listen to YC's a- advice and, and actually ship as early as you possibly can.

Matt 32:20 I think knowing when to carry on as well is really challenging.

Nicolas Dessaigne 32:23 Mm.

Matt 32:23 So like we, you know, version one, um, like there were some signs that there was like a real problem there.

Nicolas Dessaigne 32:31 Mm.

Matt 32:31 Um, and then with like version two of the product, there was, there were more signs, but we didn't have product market fit. And I mean, it was two years-

Nicolas Dessaigne 32:38 Mm

Matt 32:39 ... uh, before we had product market fit, but we kind of kept going, um, 'cause I think there was just enough like hope that there was like a re- there was a real problem here, but we just couldn't, couldn't quite solve it. And I think it's very hard to call like when should we like give up and pivot?

Nicolas Dessaigne 32:53 Where did that intuition come from?

Matt 32:55 I think we, we had had the problem ourselves-

Nicolas Dessaigne 32:58 Yeah

Matt 32:58 ... before, and so we had felt the pain personally. Um, so we, deep down, we, we really felt like it was a worthwhile problem to solve.

Eric 33:09 Yeah, and we sort of knew we like weren't, weren't maybe quite there yet.

Nicolas Dessaigne 33:13 Yeah.

Eric 33:13 And, but we knew like it w- there's something here, and we, we knew it because like, yeah, we had felt the, the pain for all those years building other things.

Matt 33:22 But being really, really close to customers I think is, is the other piece of advice. Again, it's like classic YC advice. But, um, like if you are talking to them every day and getting their feedback, you will, you will realize things that you would never have realized like in, in a vacuum on your, on your own.

Nicolas Dessaigne 33:40 Guys, I really enjoyed this. Thanks for spending some time with me. That's it.

Eric 33:43 Thanks for having us.

Matt 33:43 Thanks.

Nicolas Dessaigne 33:44 Cheers. [upbeat music]