David Cramer of Sentry on never going upmarket

Hey, thanks for having me on.

I am good. I'm in a perpetual state of as I always am, I would say.

It's interesting because sometimes I do very little. I'm not shy about this, by the way. Sometimes I'll have weeks where I put in very few hours in terms of hard work, and then other weeks it'll be like 60 hours or something. Recently it's been a lot more, but it kind of depends because it's not really any different than any other person. If you have something that's really interesting and feels high value, you kind of want to complete the project or you want to do it. And as a founder, there's this inherent you can see the ROI of the thing you're doing. And so if the thing is like, oh, I have a bunch of emails to respond to or a bunch of meetings to go to, I'm like, it's not that useful. And so I'm just not going to do those things. And so it kind of depends. But yeah, it goes through cycles. But recently it's been a lot busier.

Yeah, definitely by choice. I tell people at the company even that I am still doing this truly by choice but also even Sentry because I've hired the CEO and recently the CTO and so I basically stepped out of both roles that I am at the company because I think my equity value effectively grows faster with me at the company than without me at the company. And so I think it's more lucrative for me to be there but also I've done this most of my career at this point. So it's a little bit more than that.

Yeah, it is quite complicated. I think a lot of people cannot do it for what it's worth. I think I struggle with it and I think you got to find the right people and it doesn't mean I agree with everything we do at the company and I think we get in healthy debates a lot of times at the highest levels about like is this the right thing? Should we be doing something else? And I think the most important thing we have is that we all recognize we are trying to do well for the business. I think particularly especially with the CEO and the CTO roles because executive hiring is hard in general. I think we found people that really wanted to succeed and the CTO because the CTO in our case is effectively also driving the product. It was really important for me. I wanted somebody that was kind of like me that would make similar decisions to me. I didn't want to take risk. I think we nailed that. I would say our CTO is a lot like me. They're less in the weeds, but they are much sharper in terms of academic background and things like that. And so it's hard. You kind of want to get to a point where 90% of the decisions they make would be the decisions you would have made. And then you got to accept that some of them wouldn't be and that's okay. And sometimes those will be better decisions, sometimes it'll be worse, sometimes it's neutral. Most of the time it's neutral. It doesn't matter. But it's just one of those things. And I think a lot of people struggle with that. And as soon as you forget about the illusion of control because control actually is not that important most of the time or rather the idea of a title or something being how you have control, like I have immense control at the company. I don't need a title to have influence. Influence is probably a better word than control. And I think as soon as you let go of that, life gets a lot easier and I let go of that a long time ago.

And early on at some point everybody's like, 'Oh, you got to hire people and let them do the stuff.' That is absolutely the wrong advice for what it's worth. But it's kind of like the company gets big and you've got to do the thing that is most important for you to do that other people can't do. Especially as the sort of there's usually somebody that is like it's a CEO but it's also the CEO founder is like a specific persona. It's like the buck stops with you kind of thing. So work has to get done and so you're doing that and I'm like I'd love to write code but I have the other stuff has to happen you know and more recently I've actually had time to build stuff and it's been fun. I also block off two-thirds of my calendar and so you just can't schedule stuff and so it's actually there so I can work on something or experiment with something but I will say that's the ambition. It still ends up being I can't really contribute to the main application that easily because it takes a lot more sort of focus time than I have.

It's a thing I'm most proud of. It's the singular goal I've had in the company is like it should be everpresent. It should dominate the industry, but it should do it because we make the right decisions. We do what's best for the customer. We don't try to cheat people or anything. Build a good product and then we make all the decisions around that to support that kind of thing. And it's at this point it's great because it's worked. Takes a long time to get here obviously and it's been really good. And actually that 4 million number I don't know how old that is. I also don't even know how we calculate that. It's such a it's hard to calculate those numbers obviously the team count is much easier but it's basically used by everybody. It's kind of like the default which is cool. It's really hard to achieve that but it's fun because we get to do big brand marketing and stuff and people kind of know what Sentry is. And so now we do stuff and our current agenda is like take something that's frankly just boring. Sentry is not an interesting company and just make it fun. And so, you know, we put out ridiculous merch and stuff like this and yeah, it's the billboards and whatever marketing campaigns we have is really fun.

The theme was well this started like years ago when we did our first Billboard and I'm like I just want to do something that's shock and awe because I think the goal of everything is attention and I'm like well let's just do something that's so preposterous that it doesn't make sense and then double down on it that it has nothing to do with Sentry so it makes even less sense and so it was like the first concept of this was called 'Sentry can't fix this' that was the campaign and it was a minotaur playing Beat Saber and knocking things off shelves. You couldn't really visualize it because line art, so it didn't do the illustration. Probably we could have done a better job, but it absolutely makes no sense whatsoever. And that was like the first advertising campaign we did. And the current one is supposed to be the concept because the creative team tries to connect it to reality is we're like production monitoring and you look at dashboards and numbers and the data is useless most of the time. It's like cool, you have all this stuff, but it doesn't actually make a lot of sense. It's like why do we have all this? It's just a bunch of noise. And so the campaign was like make it make sense kind of thing and we're just like let's just throw random things that are so obscure at them and then it just turned into like oh here's these hideous face things which we've kind of done something similar to that before and it's just fun you know.

The way I always articulate Sentry is like we just help you debug problems. It's kind of simple and it's like or bugs and bug is a loose definition. Like what do we think of as a bug is like users not developers because like you talk to a developer about what's a bug and they're like well that's not really a bug and they have some annoying definition right and you're like no it's broken it doesn't work right and so we think of it like it doesn't work right that's a bug and so the idea with Sentry originally was like well LLMs are really good at pattern matching you can actually simplify how you reason about bugs like explain them better and so it does that pretty well actually basically take all this noise that is the telemetry and make it make more sense so that's like so it does a bunch of stuff and then it can propose patches basically. And the patch generation is about as good as everything is. So it works sometimes, it doesn't work other times. And our sort of hypothesis is, and the tech isn't here yet. And to be fair, frontier models are what we're banking on here is like if we could automatically probably not like merge PRs, but if you could imagine that you have a queue of proposed fixes that are more or less correct, that's like the most trivial 10% of your bugs that you're just not going to fix because they don't matter that much, but they're easy fixes that are just ready to go. I would love that. And I would actually pay a bunch of I'd pay like $200 per bug fix or something. I'm like, we will have that at some point I think in the next year or two, but I think models have to get a little bit better before that to become a real possibility because no matter what you try today, it's not quite there. But you can still do some cool things today.

Yeah so I like building things and I spent a lot of time just on personal projects and whatnot and I built our MCP server and part of this is like I have a lot of domain expertise. So, I would not consider this luck or anything. I have a pretty good sense of what is going to be useful and what isn't. Doesn't mean I know everything, but if I think something's going to be useful, it's probably going to be useful. I don't necessarily know how big it'll be, but when I saw MCP, I'm like very clearly this is a way for me to plug Sentry into a workflow that is very obviously valuable. And it's like I'm sitting here in cloud code or cursor or something. I'm working on my application. I'm building new features or I'm fixing bugs. Some of those bugs are in Sentry. I want to pull in that context. I could copy paste it or I could just copy paste a URL and have it automatically pull in the best version of everything and I could curate the best version of everything because I know better than the customer because I am actually the domain expert. And so I'm like MCP lets me do that. And so I saw that I'm like and they had started doing the authentication protocol and I'm like this is the time when I can make this possible. And so I built that pretty quickly and prototyped it and we've still been iterating on it and trying some new novel techniques on it and it works unbelievably well. So spent a lot of time on that. And I certainly dog food a lot of our tech and coerce a lot of our team to dog food on my personal projects as well.

Yeah. I think it's useful because it's a different perspective because when you work at Sentry, we might as well be an enterprise company. It's a very large application at a very large scale. You can't ship quickly. The consequences of breaking things are serious. The scale is serious. I have a bunch of zero user apps effectively. Some that are super overengineered, some that use different technology stacks and so it gives very different perspectives so it's useful and some other people at work have similar things and so I think the value's there but yeah and so we spend a lot of time dog fooding a lot of this stuff but I think the other interesting challenge and you know early stage versus late stage companies or startups even have this is like we're not early stage we don't take blind risks you know we can't be like yeah let's bet the world on some wacky idea that's definitely not probable you know and so for us even like Sentry and MCP is like, 'No, these seem plausible outcomes. These seem likely to have strategic value no matter what world comes to be, you know, kind of thing.' Versus like, take Vibe Coding. I'm like, if you look at say rewind time a little bit, because lovable and bolt are a little bit, we know a lot more now than when they came out. I'm like, this doesn't make any sense. Like, there's no world where we're just generating an app from a sentence or something. It just doesn't make any sense. It's not possible. And it's much more obvious now to more people, but we would never have bet the future on that kind of thing. But as a startup, you have nothing to lose. You can do those things. And so we have a little bit different of a strategy where we're always like what is valuable no matter what happens kind of thing. And so that's always incremental from there. And so we spend a lot of time thinking about that. And Sentry is the same, MCP is the same, but then also it's like well what do we control and what don't we control and then how do we be defensible and all these other little things that go in there. But to go back to your other question of like I've not written code by hand in enough months that I've lost count, but it's six to nine months at this point exclusively via cloud code or cursor. Pretty effective at it at this point. A lot of times it's inefficient, but I can get a lot done now. Like the MCP server is entirely generated at this point. I wrote it by hand originally, but everything is generated plans pretty well. And so I would consider myself probably forefront expert of how this stuff actually works in reality versus loud opinions on the internet which there are a lot of those.

I think first and foremost like people latch on to this idea of new things way too often like the models got to a point and I think it was like sonnet 3.5 or whatever and then they were good enough and that was it and they've gotten a little bit better but it's very incremental now and so everybody had this idea of like oh the new one's so much better and everybody latches on to this new model syndrome it stopped it's done like they are better but like even the codex versus versus pick any Claude model it doesn't actually matter I can use any of them and be just as effective I can't even tell you the difference because it's so insignificant to me. Now, I am sure there's tasks that matter and don't matter and people will go through all these almost like voodoo. It's like, oh, you got to do this and then clear your history to have a new context window. And I'm like, doesn't actually matter most of the time. And there are probably things that are useful here and not useful here, but I can run compactions over and over and be fine. I can run new context windows and be fine. And so, I think you just got to figure out the workflow. That's all it is. It's a workflow. It's which tool works best for you, which approach to this works best. I will argue that like purely agentic, I don't know if you'd call vibe coding because I still like I spent a lot of time with systems engineering in terms of sort of defining how I needed to look and then going back and if it's not like that, I sort of reprompt it over and over. I think cursor is the best tool. I don't use it as much anymore, but I think it's the best tool to do this because you can still kind of review it as it goes, which is a really good workflow for the real world. But again, use the best tool for you. But I think a lot of it is like you have to provide the right constraints. And I think the hardest part is you need it to basically always look up information. Like you can never just be like ask it to do something because it doesn't know anything. And I don't know why people don't understand this. You can't train the models. Like it has no it will never be smart. It has to have real-time information. And to get real time information has to access the internet. So I'm basically whenever I'm prompting it, I'm like go look up the information online. 100% of the time that's what I wanted to do like look up the latest versions, look up the docs or do the thing from the shell, you know, and so I think trying to force it to do those things is really good. And there's simple things that most people figured out at this point like definitely use planning mode if you haven't have it generate like specs or something because those are pretty useful. You can review and they're also like a good way to resume work and whatnot. But there's not a lot of magic. A lot of people again they'll come up voodoo and stuff. It literally doesn't matter. You'll come up with similar results. You just have to have a workflow that works for you.

I kind of think so, but it's always hard to say because the image stuff is really cool right now. Like nano nano banana what a ridiculous banana pro. Very good. Very very good. It's phenomenal. And I love it because it's like those are things. It's pretty easy to write code, don't get me wrong. It's pretty hard to come up with lots of images of random things and especially those kinds of things and so like what I want is things that fuel creativity and you cannot like image generation is a creative endeavor at the end of the day. Code generation is not a creative endeavor. It kind of needs to be deterministic and so I want to see more of that or like the world labs stuff is really exciting. I don't know if you've seen that but it like look up world labs. Phenomenal like what the text's doing and it basically like I don't know how the I'm not smart enough to understand this but it'll generate like a 3D world of an image very cool and so interstellar style kind of thing. I think kind of yeah but like you can imagine like if you took this and all of a sudden you could walk around in a full world or something just from a snapshot and so there's going to be a lot of really cool things that'll come. I love it. Phenomenal pro tip: you have a significant other or family member you want to impress them, go to ChatGPT and ask it to generate a song with lyrics about them or embarrassing facts and be like, 'I need lyrics to a song.' And give it the genre of music you want. Plug that into Suno. It's phenomenal.

Oh my god, it's so funny. You should go to David's Twitter page and look it up. It's and the Enterprise ready hat. Shout out to Michael Green. He was just on this podcast as well. It's just so funny to me. One of the guys at Sentry over the weekend, the holiday weekend, he built like a mobile app. It's this is not public probably because it would empty his bank account. You can basically take like a selfie and then randomly generate a bunch of nano banana prompts. One of them is the it girl selfie. And so that's what started it. And then you can obviously upload photos to do the same thing. So I'm just like, cool, let's go.

There's a bunch of others obviously, but you can almost like this will be the new TikTok generation or not TikTok, Snapchat, like the filters and everything. And it's super fun.