Allie Howe: Welcome to today's episode of the Insecure Agents Podcast. I'm your host Allie Howe and today I'm here with James Cowling, CTO and Co founder of Convex. James, do you want to introduce yourself?

James Cowling: Hi Allie, yeah, it's nice to be on the podcast and also welcome to the the Convex office. As you can see the the sign behind us for folks watching the video.

Allie Howe: Yeah, great to have you here. We we met at an event recently talking about vibe coding and the kind of guardrails needed to ensure the future generation of software engineers are safe, I suppose, and I think really clicked in terms of what the industry needs. So yeah, I'm excited to chat more about things. Awesome. Yeah, how old is convex?

James Cowling: Yeah. So convex is is ancient by startup in startup years. Convex is about 4 1/2 years old and we're small. Convex is 20 ish person company. Now we're not small in terms of presence. There's, I don't know, well over a million projects running on Convex. But yeah, we're, we're, we're 4 1/2 years in. And partly, you know why it takes time to build something like Convex? It's an infrastructure company, it's a database company. And so, you know, the first few years, they're just building a very complex thing, which is a distributed database and, and, and an execution platform. And then once you have that there, then there's the battle of, of trust.

And I think it's a very reasonable question for folks to ask, should I be depending entirely on a, on a startup, right. And so early on, I think it, it took time for the community for the industry to be like, oh, wait a second, this this convex thing, I've heard about that before. It makes sense, right? And now with there's been a inflection for us and growth is kind of out of control right now. There's one of those things that what I said phrase slowly then all at once or whatever. Whereas for the longest time, yeah, it's out there kind of doing reps just kind of getting people to be more familiar with the platform.

And now we're just seeing a massive inflection and things are really, things are really moving right now.

Allie Howe: That's amazing. Yeah. I feel like infra startups are really well positioned for this new like AI native world where like AI needs all these new like primitives for running. I think we're seeing that with the cloud right now too, where we've got different like sandbox companies like spinning up because like AI agents can't run on the cloud as well as traditional code can. Also like auth is being rewritten for like context aware auth. So yeah, I feel like infra startups are a huge part of what's going on right now. Really well positioned and really excited that Convex seems to be in the middle of all of that.

James Cowling: Yeah, I, I kind of have this theory of this kind of bimodal distribution of, of, of companies in, in the tech industry right now. You have the kind of platforms and then you have the agents, apps, everything going on top of that. And I think they are have very different demands and they should look quite different. I think it's more than ever before that these companies look quite different because there are extraordinarily fast moving product companies right now that should be moving very fast. I will, I'm sure we'll talk about this later on. I can't go to them and tell them to care more about security or take care more about architecture.

They don't care. They want to move fast. They have to move fast to survive, right? So there's kind of this very quick moving rather heuristic companies just getting things good enough to work that depend for their existence on platforms with very strong guarantees that are very dependable and building blocks that kind of help enable them. So I'm on the I'm on the platform side, right, which now that still means we're a very fast moving company right now because the industry is out of control right now. I'm stressed every day. I mean, a good stress, right? But like my head's spinning every day. But sometimes I think, wow, I'm so glad I have an infra startup because we can do a little bit more long term planning and we have a kind of a durable set of core building blocks that people build on and those evolve over time.

They're not just changing every day. And, but, but I, I'm excited. I, I don't think that that we're going down this path where quality is dead and good abstractions of building blocks dead. To the contrary, I, I think they're more important than ever before. And we're seeing that agents benefit tremendously from having good building blocks to work with. But there's just these two different kind of communities of builders right now, and I think that's largely a good thing.

Allie Howe: Yes, definitely. I feel like there's a lot of vibe coders out there that are just like, I just need to ship fast, get something to work and they don't really know or understand for the benefits or trade-offs of choosing different platforms or different building blocks. One thing I see commonly with vibe coders that maybe don't have an engineering background, they're just doing this kind of like for the first time, they'll just spin up any like sort of database platform that they've heard of and lot to like throw anybody like on the bus. But there's a popular one where people typically run into row level security role issues where they don't configure those because that's not really like a secure by default feature of that platform.

And so they ended up like shooting themselves in the foot usually with that platform. So having the correct building block is definitely key for developers.

James Cowling: Yeah, so I mean, I, I've been in the industry for a long time now, you know, 20 odd years and I have been fortunate to work on some very large systems at Dropbox. I led a lot of the large infrastructure projects like moving Dropbox off of S3 and building a million no distributed storage system or building a 4 million transaction per second database. It's very large systems with large numbers of engineers and having been very fortunate to work at very large scale and then also working at large scale and trying to build systems that are not just, you know, scale in terms of number of users.

That's when people say scale, I think into the number of users. But when I think of scale, I really think about just does will your code last? Will it be maintainable over the long term? Is it going to be a disaster one day? Is, is complexity going to emerge to the point that it all grinds to a halt, right? What's the difference between a, a prototype demo and a real company that pays the bills and you can feed your family with, right? And not just for a week, but for for years, right? And I think any senior experienced systems developer will tell you this, that the, the core differentiator 1 is simplicity, simple system scale and work.

And that there are just some kind of building blocks, primitives that are right and some that, that are, that are left designed for, you know. And so I think sometimes like you can see an app on the surface level and it, it looks like it works right. But is it built from these kind of core building blocks that compose and and, and, and stay simple over time? And so I would argue that some of the, some of the primitives that are really needed for apps to work over a long time is stuff like transactions. You know, your code should do everything in a transaction. So there's no race conditions because reasoning that race conditions starts falling apart at scale and even Claude Opus whatever is going to have troubles race conditions if you don't have transactions in your code, right?

I think stuff like abstraction and modularity, hiding stuff behind functions, right. You mentioned row level security rules. That's a tricky one too, right? So, so there if there's a, there's a very, very popular design pattern amongst web development. I get why it's popular, right? Which is to expose your database to the to the world. Have your all your business logic on your client, on your client app, have them issue, say, SQL queries against the server and then use configuration files to say you know which, which rows can, can, can client code read. That's very convenient because your client code has unbridled access to the database unless you've locked it down.

Now it turns out it's pretty hard to get these rules right. And LLMs do great at type safe code where it's correct or it's not correct. LLMs do really well when there's only one kind of correct way of you built it in a compiles, you know it works, right? The challenge with config files like row level security rules, it's very, very hard to know whether they're correct. And yes, it's true, there have been a number of incidents where people have accidentally exposed their data because it's like complex configurations. But at the end of the day, I'm a big believer in architectural primitives. I don't expect application developers to care about that.

I'm not going to go to people and say hey, use convex because it's more consistent or use convex because blah, blah blah, it has transact like I don't expect someone to care. So what I want to tell people is, hey, just use Convex or a platform like Convex because you have found that your agent produces systems that just happen to work and are really simple and don't degrade over time. And I think that's what we're going to see. I mean, that's why Convex has experienced a lot of growth right now is people are finding, oh wow, if I just use this platform with building blocks that are hard to misuse, the agent just does the right thing.

It's basically, I think about an agent as a junior engineer, as a, as an intern. It's kind of that's the, it's, it's, it's like working with a really, really productive intern. How do you manage an intern? Well, it's, it's by creating the right building blocks, the right guardrails, the right primitives that are hard to use wrong, don't leave a loaded gun on the table. I think that's the industry's going to really shift towards this way of thinking, right? Where you have agents just running wild. I don't think we can slow this down, right. But then like a baseline, a very safe baseline upon which they can play.

Allie Howe: Yes, now that's amazing. It's like, how do we take this, which is so powerful and it's like a junior intern but maybe doesn't know the most about like security or or is not the best engineer, but how do we put boundaries around that? How do we put like hard boundaries around that? What can it access? And can it access the full database or just part of it? So really interested to chat too about like the security side of Convex, how Convex can help developers ship securely by default, which I feel like it's such a huge problem in our industry right now where people are looking to agent development and saying, hey, like, how can I actually like ship secure by default and not have to overthink and come up with all of these like rules, platforms, all of that myself?

James Cowling: Yeah. The first thing I'll say is you can write insecure code on Convex. You can write buggy code on Convex. It's just harder to do so, right? So I'm not going to promise anyone that if you use Convex, your code will be secure, right? But what I can say is my philosophy on how to build something that's secure by default is to start with your database is invisible to the world. That's your private state. Not only is it not leaked to the world in terms of access, from my perspective, it's also not leaked to the world in terms of API or abstraction. The the client code does not need to know how your database is laid out.

And that's actually really nice long term because it means you can refactor it and you can make changes later on, right? So by default nothing is available to the outside world. Then you write in Convex, you write query and mutation functions. These are PC RPC services, right? And you expose RPCs and so you write some code, you write the code might be called, get user information, right? And so all you call the RPC, it looks up the user state, it returns the state. The only thing that function can do is return the thing that the agents asked it to return, right? So it's an opt in model.

Now, once you add auth to your application, hopefully the agent or whoever's writing the code, well, the first thing you should do in that function is, is check if the user's authenticated to do so. I was, you know, vibe coding something yesterday. And at some point I'm like, hey, just just like double check that this app is secure. And I was like, oh, I actually the agent was like, oh, I, I am not checking auth. And it went and added all these auth checks, but it was very easy for it to do so, right? It's very easy to look at it. It's very easy to, especially for agents to perform local reasoning, right?

Here's a function. It's the get user state function. Look at that function. Doesn't check the auth, doesn't expose only the things I want to expose. It's much harder to reason globally. In fact, LLMs are very, very bad at reasoning globally, right? Does this configuration adequately lock down my entire system? So, you know, I think I think a lot in terms of simplicity, compartmentalization. How can we reason locally? You know, transactions are helpful for that. If, if this code is transactional, I don't have to worry about what happens if a user is added at the same time as some user states getting read because they, they run atomically and independently.

Same thing as security, right? In Convex, everything is, is everything's inaccessible by design until a function is added to expose that state. And you can review the security of that function just by looking at that one function.

Allie Howe: So then at that point it seems like most of the security simulations are around. Like, is the LLM following the user's permissions? Yeah. And is it is it following kind of good developmental practices?

James Cowling: Now I have a lot of faith the LLM's will get better at this. I have a lot of faith, especially I may have a lot of faith that you can you can write a prompt that says, even if they they didn't do this by default, you can write a prompt that says, hey, never have a public your mutation without checking that the users authenticated to do that and and Claude will figure it out. So I feel pretty good about starting with these building blocks and asking an agent to do the good thing with those building blocks. I feel pretty skeptical about asking an agent to secure a whole system.

Allie Howe: Yeah, I do too for sure. But that's awesome. You have like auth building blocks. Is there any other security like building blocks that come like with Convex that help?

James Cowling: Let's let's think about it. I mean, primarily I think about it in terms of auth now, obviously, like if a pro accounts, we have all sorts of, you know, compliance and all sorts of business things in the background, right? But but the, the, the what I care right now is, is building blocks and the basics for users who don't care. Maybe we can talk a bit more about that because this is this is the era. This is the era of just not caring, which is fine, right? But the industry is moving so fast. The average start up just does not care about security, does not care about correctness if it comes at the expense of their development velocity.

So for me that the big challenge is how can we add stuff like authentication, for example, in a way that's very, very ergonomic and doesn't slow them down because people, you know, people aren't even reading their code anymore. That's, I mean, maybe I sound like sound, sound like it like it like an old man, just like being amazed by this now, especially if you're watching this video in in a year and you're like, of course, James, no one's reading their code. But I will say that I think amongst like professional developers, it's happened almost in the past weeks, months that it's become almost accepted that you just, yeah, you might not even read the code that AI is producing.

Allie Howe: Yes, we definitely have like a code review problem. And I was super interested to hear you talk about that at Daytona's event where you presented and talked about how, yeah, very few people are actually reading their code nowadays. And there's also lots of people that are not engineers or not engineers by trade that are developing agents for the first time. And the difference between a proof of concept agent and one that's ready for production, one that's going to last and like feed your family for years or whatever is, is huge. And how do we review code and keep up with the speed development?

James Cowling: It's an arms race right now. It's an arms race because of the incentives. I'll do some fun name dropping. Right? So we had lunch yesterday with Pete from Open Claw, also previously known as Claude bot and then Mold bot and now open claw. So if you're not familiar with what this thing is as of as of right now, when, when this podcast is getting recorded, it's the hottest thing on on Twitter, right? It's the hottest thing in agent development. So open claw is is actually very cool, right? It's the ability to run an agent locally on your computer. Some people are buying Mac minis to run it, you know, but then they're pretty quickly giving a lot of credentials to that agent, right.

And then you can text your agent via WhatsApp or whatever and say, I don't know, book me flights, do whatever you could like there's, there's A and so why, why am I involved in this story at all? Claw hub runs on convex Claw hub is the repository of of skills. So open open claw has a whole bunch of skills you can give it and some of those skills like talk to Spotify and set up whatever, you know, it's just a bunch of it's like, it's like your agent is working for you and it just, it's extensible and it's all these skills you can add. So what's interesting about this?

Well, one wow, what a terrifying world we're in from a security perspective. Yes, for. Sure, right. Like, Oh my God. And like, and everyone who uses this knows it's kind of sketchy. I mean, like, he's not sketchy, but everyone knows wow, this is a bit scary, isn't it? Like, wow, I have an agent and it has access to do a whole bunch of stuff and it has a bunch of my credentials because I wanted to do cool things. And this feels pretty risky. And there's a lot of skills on Claw Hub which are like nefarious, right?

Allie Howe: Yes, there's a what would Elon do skill. Like did you see the guy he hacked that Claw Hub, whatever the skills registry is to increment the counter so that his skill was the most downloaded because that's what people are establishing trust right now. That's the only metric to go off of the most downloaded. And so like this skill would ask what would Elon do? And it looked totally fine the skill MD but there was a hidden logic dot MD file in the skill directory and it called out to his server and would like send your host name there.

James Cowling: Oh yeah, it does not surprise me at all, right? And because there's an open repository of skills, right? So wow, where this is Wild West we've never seen before. I don't think in the takings. I can't personally recall a time where it's felt this Wild West. But I'm not here to moralize. It's like I'm not here to say that's wrong, right, because people are using it because it's so cool and it's so powerful. I can't tell someone, Hey, don't use this thing. It's a little bit insecure. Pete, by the way, very smart guy. That's that was such an interesting conversation because this often times we have non engineers vibe coding something on convex and they might build an app that gets unexpectedly popular and maybe some parts of it aren't efficient or whatever.

And but they just don't know the difference between using an index and performing a filter. And then we go back to the team and say, Hey, you need to make the lint rules better so the agent knows to do the right thing. Normally that's the conversation. The conversation with Pete. He's like he is a very experienced professional engineer, right. He knows claw hub is just vibe is just the vibe coded thing. I don't think you looked at the source code right? Runs on convex, so as a result mostly works and mostly scales right. But it is somewhat inefficient the way at least it was built originally and he doesn't care.

And that was such shocking. Such cool realization, right? Here is a guy who knows how to do it by hand and he knows that efficiency matters, blah, blah, blah, blah, blah. And but I can't, I can't tell him, hey, Pete, like change this thing and your convex bill will be lower. He's like, I don't care. It's it's running well enough right now. It's. You know, obviously he's like in the spotlight right now. He's busy, right? He doesn't like the, the, the incentive to go and care about the details is kind of going away to some extent while there's such an incredible growth and momentum in the industry.

So then it kind of really turned the turn the tables around on me because I was like, huh, I, I can't tell this guy to go improve his app now, but Pete's very happy with his convex bill. So great. But it could be lower, his bill could be lower, you know what I mean? 1400 was the original. Yeah, yeah, and like it sure. And by the way, it's it's a super popular app, right. So sure, he built as far as he's concerned, that's that's nothing compared to this is like the in the zeitgeist massively popular app that's that he vibe together as running on Convex is great.

So it's a success story for them, for him and for us. But I can't, yeah, there's there's no longer a dialogue where I can say, hey, hey, dude, you need to care about this. I don't care. That's why don't you fix it. And I was like, wow, that's that's such an interesting situation. Like I now I'm more I or the convex is responsible for the quality of this system. So it's almost like the, the party's here, the convex team. And, you know, I'm saying it from the convex perspective. The same goes for other platforms. The parties involved are the convex team who build the platform, have the building blocks, and have the right incentives for agents.

You know, like maybe some feedback to agents that says, hey, this query is inexpensive, why don't you make some changes to it, right? Then the agent then builds the code and then the developer who is like the product lead and like our customer in many aspects is the agent now. Right. So true. Yeah. And so I can't say yeah, the days are over where I say, hey, you should care about adding index. No, like Convex has to tell the agent firstly we should make it even as of right now we're already I think the most kind of guardrail centric platform out there, right. So Convex is already like really tight in terms of it's pretty hard to use Convex wrong, but there are some ways you can use Convex wrong.

Still, there's some inelegances in our platform and the pressures on us now to remove the parts of the our stack that are inelegant or have feedback to the agents other via ES lint rules to say, hey, I noticed you're reading date inside a query. Maybe you don't want to update that query every time the date advances. That was actually one thing that was in in cloud or probably what we have to do. And we had this conversation yesterday. We probably have to add some functionality where the agent can ask Convex, hey, how's my app been performing and Convex can say, hey, guess what, this function is reading a lot of data, maybe you should reconsider how it's architected and then it can go back and re architect the code to do so.

So it's in a strange world where like the product developer is responsible for the interaction, the quality, the marketing, those are the quality of the experience, the marketing, etcetera. And the platforms are almost platforms in collaboration with agents are responsible for the quality of the software and I don't see that changing.

Allie Howe: Yeah, I know. That makes sense. How do you, if all that amazing interaction is going on behind the scenes or under the hood between Convex and the agent of saying, hey, like you know this, if you change this, it would cost you less or it would be more secure if you did this. It's a lot of value add that maybe the user isn't seeing. So how do you make sure the user understands Convex is doing a lot for you? Like this is like something you want to use? Like where's the go to market sales? Thing

James Cowling: yeah, it's hard well so so 1 so so for example, one anecdote is Jamie my Co founder who you met this morning CEO of Convex very busy guy he literally patched claw hub and pushed it to the GitHub repo to lower that bill because there was something there was a couple of things about the app they were just inefficient at very large scale. Most people don't have to care about that stuff but Claw Hub that's massive scale right so we care we don't want people getting ripped off by us. We don't want people doing inefficient things on Convex it's not good for the platform. We want long term satisfaction, right?

So, so I'm not, I don't really care about like extracting every cent out of a user right now, right now. I don't ever want to extract every cent out of a user. I want them to like have a good experience. Now it's hard for me to explain to someone. It's hard for me to say like, hey, because I think a lot of developers even listening to this might be like, like maybe not believing what I'm saying, like unless you've spent time building things of scale, maybe you're like, why do transactions matter? You know, I've had conversations with like, and it's, it's always with people who are 23 years old, right?

But they're like, well, no, Claude can can totally handle race conditions and you don't need transactions because you don't understand the agents going to solve this problem. And I just don't agree. Right. And I I don't agree and I don't know that I can convince someone. I mean, it's it's it's curious that it's mostly like inexperienced people or non developers who have this takes even folks who work at Anthropic and OpenAI. Many of my friends are at these companies. They will readily agree that their models aren't great at systems code right now. There's a reason for this right? System code often times involves either non local reasoning.

How does one this this this this system over here interact with some other part of the code base right? That maybe isn't even the same repo because it's you know, from a different service. And they often involve emerging patterns over time. I built something today. How does it work in three years time when all these architecture features have been added to it? And so For these reasons, and also if they involve skills that are pretty rare, there's not that many experienced systems engineers out there. There's not a lot of good training sets. So as of right now, no matter what model, you have not very good at distributed systems systems code, right?

But anyway, that aside, I'm appealing to authority here, right? You might not believe me and that's OK, but so I have to impress you, you the developer, by virtue of the fact that Convex can build an app that's not possible in any in any other platform. And I think we do a pretty good job with that right now. Now this, I would say auth is still not great in Convex. Assume we still haven't nailed it. I was building an app yesterday. We yesterday we had this build on Convex day. So it was like vibe coding all day and my brain hurt at the end of the day.

And there was this little bit of a bootstrapping problem. It was a little bit frustrating to get the app working and log into some dashboards and stuff and get auth to set up be set up. Once that was done, I was my mind was blown and I I shouldn't have been like maybe I should know this already, but I just could not believe how well I was using Sonnet 4.6 as a new one. I think 4.6. I just could not believe how well it performed. com it just for the rest of the day basically did not make a mistake. And so I'm not in this kind of eat your vegetables mindset where I'm like, use use convex or you because it's it's better or more secure.

I'm just I'm like, hey, just try it. I think you will find your agent is incredibly productive on the platform.

Allie Howe: Yeah, It'd be interesting to see if you made the exact same agent but with different building blocks, like convex versus something else, and that was the only thing that was different. And then you could like use that agent overtime and show like a dashboard of metrics around like how that actually improved. Overtime,

James Cowling: maybe we should do that, but I would encourage folks to try that. Yeah, I mean like I'm not going to hard sell anyone. Like I would say I'd love for folks to try Convex. There's a few little concepts to wrap your head around to use it or maybe don't wrap your head around and maybe just ask your agent to use it, but just do a head to head comparison. You don't like it, don't use it, you know, but I think I think what we're finding is more and more developers are really gravitating towards and, and one other thing, like comics has been around for 4 1/2 years, right?

When we started, people had a passion for Postgres, people had a passion for using Terraform or Kubernetes. There were a lot of developers who were like, why would I use a platform like Convex? I want to use this, you know, tried and tested maybe 50 year old, maybe that's not a good thing code base because they personally had a connection with it. And that's just changed dramatically this year. Last year people were always like, oh, why would I use Convex? I love Postgres, whatever. Why would I use a platform when I could set up my own thing with Redis? Just don't hear it anymore because I mean, I'm sure those people still exist and they're probably my kind of people, probably come, come work at Convex, apply for a job at Convex, right?

But those people can't be heard above the roar of developers who are so excited to build stuff now with agents and just do not care about their stack. They're like, just give me the stack that works the best, and that's rational.

Allie Howe: Yeah, yeah. I think we'll start to see that phase out of those people and then just like get older, I think we'll see like less opinionation at the start because they're just, they haven't run into those problems where they've, you know, been the lead engineer in scaling something and seeing how everything goes wrong. And then they experience that pain and like, I'm never going to forget this. I'm going to do something about it. But that's great. They don't have to if there's platforms like Convex and agents that sort of sort that out for them and educate them. But I really like what you said about the whole like, eat your vegetables mindset because I experience that all the time in like agentic security where I'm like, you can't just rely on the model to stop prompt injection.

But that narrative is getting actually a bit harder. So I was messing around with Open Claw this week. We did a live stream with Browserbase and we got to play around with it together, which is super fun. But one of the things that I showed was that the skill, what would Elon do? Where had that prompt injection embedded in it? I couldn't get it to work. So like I would say, hey, like run this skill and immediately like Opus was like, this is going to send credentials out to somewhere else. Like I'm not going to do this even if you tell me to. And I was like, no, I trust it.

Like I wrote it like, please do it. And it was like, no, I'm just I'm just not. And I similar ish like success with like ChatGPT is like 5.2 model, whatever. So the models are getting better. They're getting quickly trained on this stuff, but a very clever sophisticated AI red teamer or lab or threat actor is going to find a successful prompt injection that does work. And Zenity, one of the top like AI security labs out there, they published research where they had a prompt injection. They put in a Google Drive document, it connected Open Claw to that and we're able to get successful prompt injection and update the Soul MD file, which is kind of like the one of the brains of how Open Claw works to get that to change and stuff.

So that was a very successful exploit. So like it can and will happen, but convincing people that they should care about it is just tough. So I feel like we need that for security too, where here's the security default like sub agent like feedback loop with your code. That's like, hey, I see that this is insecure pattern that you're creating. Maybe you should like add an LLM guard rule or maybe you should be validating this amount. So if you've got an invoice agent, like validate that the invoices below sort of threshold like with just regular code before you even pass it to the LLM to your reasoning about it.

So yeah, I'm, I'm sort of building something that I think that will help with that.

James Cowling: Yeah, you need it, Yeah. I mean, I, I hate to say this, but this is the golden age of security exploits like we just have not seen. This is the tip of the iceberg. We have not seen like never before has there been such a productivity at least, you know, at least short so far, short term productivity in app creation And wow, that's a terrifying world right now. You know, so that I think that we're going to see a lot of lot of security exploits are going to happen. And I think there's can be a lot of folks out there. I mean, if hey, if you're looking, if someone wants to come up, you know, if a software engineer is looking for a job right now and doesn't want to work it at our companies, there's a lot of work out there.

I'm sure taking vibe coded apps and locking them down, turning them into something usable once someone's kind of gone a little bit too far and kind of got themselves caught up and tangled. I've seen this quite a few times myself. Now someone will build an app, it'll take off, they'll be non-technical and they'll be like, I don't know what my app is doing now. I don't know why or is not working. I don't know why this thing happened and Claude can't figure it out. There's. Oh, my God, there's a lot of work for consultancies out there. But yeah, I mean, Oh my God. Security is and, and containerization will both the Daytona event.

Containerization and security are just going to be such growth areas. And the challenge is just how to get people to care.

Allie Howe: Exactly. Because it's one of those things you generally don't care about. It's like insurance, right?

James Cowling: Yeah. You don't care until it's all good, till it's not. Yeah, it's, it's, it's now somehow, you know, we buy insurance because I don't know if socially we've been trained to do it, but even like I, I didn't, I had a motorcycle accident on an uninsured motorcycle. So I'm an idiot. And then I was like, wow, I I really should have had insurance on that thing. It's too late once this happened, right? So how do we kind of socialize the idea that security matters now? Because yeah, it's you can't undo, you know, data leak once this happened. So from the Convex side, the way I'm trying to solve that problem is like trying to create building blocks that agents are able to reason about securely.

And I'm sure you have your take as well on how we can get the world to care about security.

Allie Howe: Yes. So my take on that is you have to bundle the Security benefits with performance or like something that's tied closer to revenue. So if you don't have like a compliance framework or some sort of regulatory incentive to I, I really cannot say I have to do this thing for security or else like I just can't go to market like that's a big driving factor for security. But in the lieu of that and I was doing like consulting for the past year for AI security. And so that's kind of where I found this or saw this. There's certain things that are benefits for AI security.

There are also performance benefits. So one of the things that I saw was like people were importing like MCP servers and they would come with like 100 different tools. And I would say, OK, so there's a bunch of tools we're not using here that are just sitting around as security liabilities, like sending your private data potentially if the lethal trifecta occurs where you get a prompt injection and access is not the private data. And we're going to externally communicate it through this MCP tool. And I was going to, you know, post to social media. They do all sorts of things. So it's like, if you don't need one of those tools, just get rid of it.

A lot of the different AI screen like frameworks or just, yeah, some of the like, I don't know, the the monsters managing all the AI frameworks of the world, they'll allow you to make that list like mutable where you can like get rid of some of those tools. And the LLMs perform better when they have less tools to choose from too. So that's a win on both sides.

James Cowling: Yeah, well, we already have this issue with npm, right? Because a lot of, you know, the the the current and future is, is probably, you know, web development is the primary kind of vector for application development. Primarily these are these TypeScript apps primarily pulling in dozens or hundreds of of packages from npm, npmjs. Most of the time people have no idea what's in that code. Now there's reputation systems so you own people only use packages they they know and trust. Hopefully they won't get games too much like you said. But I I had an interesting experience yesterday. I was using open code with with Claude purpose and I needed some functionality.

I forget what it was and it and it said I, they gave me a prompt I could pull in this npm package to solve this problem, but it's really large and has a lot of complexity in it. Or I could just write it in 10 lines. And I was like, huh? Now, by the way, I don't think you should hand write like you should. There's lots of things you should not hand write. You should pull in you shouldn't hand write your own SSL library, right. You should obviously pull in a tested, validated, you know, handwritten, you know, open SSL and implementation. But I thought it was a very interesting question, right?

And I actually asked the agent just to hand just to write it for me because I was like, yeah, why take a dependency on this massive package if it's just a small amount of code and there's going to be some dimension of this, like maybe we will avoid pulling in large complex base and and be building simpler versions.

Allie Howe: Yeah, I think that's happening. But yeah, there's the trade there you mentioned. It's like if this package has been out there for a while, then it's gotten eyes on it, tested, maybe it's more secure because people have iterated on it versus like I'm an LLM, I'm just going to write the code. I'm trusting that the LLM is going to do the right design patterns there. It may be a. It doesn't. And what happens when the LLMs consistently create the same insecure design pattern across all these different code bases? If I was an attacker, I might try to figure out what that is. And so now I know like all of these different apps built with this model are all vulnerable.

The same thing.

James Cowling: Yeah, I just want to like repeat the thing you said earlier because it's it's very important in case anyone's listening who wants to start a startup. And a lot of I get like my background is an infra guy and a lot of infra people think in non consumer kind of ways, right? Like people you should care because it's the right thing to do, blah, blah, blah, blah, blah, right? If you want to start a startup, anyone out there, you cannot tell someone to do something because it's the right thing to do, right? No one cares. We tried this early on. We had different ideas for a company before Convex and people were like, Oh yeah, sure, I guess you're right.

Like I believe you guys, your experience, whatever, but I don't care. You can only sell someone that someone wants and they want to pay for, and it's worth their time, especially worth their time in this frantic era where no one has time to think of anything because they're just trying to keep up. It's the challenge for security companies, right? It's a challenge for databases. You can't right now. I mean, I guess I'll be careful how I say this, but I have a huge, huge amount of respect for the Fauna DB team. And Fauna is not, I don't believe exists as a company anymore. And they had just great technology.

I think they're a cool team. I think they did an awesome thing. They were selling a more linearizable database. And I think it's very hard to market and sell that, you know, I think, I think it's very hard convex. You know, we care about these things too, but I can't really sell, hey, Convex is the end to end consistent serializable database. I don't think many developers care that much. You have to sell something that they that that that drives their their bottom line.

Allie Howe: Yes, you have to find the pain. Yeah, for sure, 100% yeah. Well, super cool. I actually, I got this head at the open claw event. Oh yes, yeah. And so I'm just seeing you guys everywhere, which I'm also seeing obviously Open Claw everywhere too. So just wanted to like circle back to that. How are you helping developers that want to try Open Claw for the first time? Is there any incentives for them to try like?

James Cowling: Yeah, I guess we're I guess we're incentivizing this, this this crazy world we're in right now. So, yeah, so there is a there is AII just found out this morning. You knew this before me. Yeah. We we have an open claw promotion. I think it's a convex dot dev slash claw. We we have six months free complex pro for people building open claw skills.

Allie Howe: Nice, very cool.

James Cowling: I guess. Wow, I know this is happening, but I just want to also just just for folks who maybe are not in the Bay Area right now, maybe kind of following the tech industry. They're in the tech industry, but they're not kind of like part of the kind of the, the, this like AI kind of hype machine that also has some real, real substance. Then at the same time, you were at the open claw event, right? Which as far as I know, I, I didn't get in right. It had hundreds of people lined up down the street.

Allie Howe: Oh, it was wild. I came from another event that was nearby, so I actually got there super early at like 4:30. And I helped volunteer and like, help them set up just since I was there early. But yeah, we had like, I don't know, like 500 people in a super small room at like Frontier Tower. A bunch of different people got to present. Peter, was there some friends of mine from Founders Inc, which I'm in that incubator program with them, they demonstrated their little robots that are running Open Claw, which was super fun. So yeah, I know it was a really great event and just it was wild to see so many people care about this project.

And I was talking to someone that made this prediction that Open Claw will be the most secure agent within the next 3-4 weeks. And I was like, why do you think? That wow. I was like, that's very much like today. And he was like, it's getting so much attention. Yeah. Like, everyone's eyes are on it. That's all. Make it better. You said your team is, like, pushing code to make comics more efficient. So all the security engineers are also, like, eyes on open Claw. And, like, I myself, I just did, like, a thing with Browserbase to speak of, like, how do I show people how to, like, use open Claw securely?

So maybe he's right.

James Cowling: Wow. I mean, look, I mean, it's just, it's so like, you know, I'm preaching to my, to my, to my infra kids out there, right? It's very easy to be cynical about what's going on right now. It's very easy to be like, Oh my God, there's so much hype in the industry. Oh my God, Like people don't care about quality anymore. And oh, that's not the way things should be done. And I like writing things by hand and deploying my own services. I like that stuff too. And I get where you're coming from. And I care about a bad quality. That's that's that's what I love is what drew me to the tech industry.

But there's just some incredible transformations happening right now. They're just happening. And there's like, if you look at the at the kind of leading indicators of like how well, I don't remember the last time hundreds of people lined up down the street in a somewhat sketchy part of Soma to go talk about a, you know, a, a tech thing. You know, I haven't seen that in a while. That wasn't the case for a long time in, in the in the valley. I've had it is it's been great for me. I've had people come to me and say, oh, wow, my friend was talking about Convex and these are not engineers saying this.

And I was like, what? When was the last time someone talked about a database with their friends, right? And so this is kind of a cool time. It's a cool time where there's excitement about about AI agents that's excited about app building. This is something about platforms again. Sure, there's tons of things that one could and should be cynical about, right? Are these secure, blah, blah, blah, blah, blah. But there's cool stuff happening. And if there's people like us who care about security or care about platforms, care about abstractions, right? The best thing they can do is try to make the the, the good, nutritious, highly secure, highly scalable functionality palatable to agents, palatable to the people who are going to be using these things.

There's a real opportunity right now. Whatever we do or say, there is going to be an incredible amount of code getting produced by agents in the next year and it's going to the industry depends on folks caring about building the right platform for agents to be able to wrangle that, you know, direction that's long term positive.

Allie Howe: Yes, I think everyone wants to be first to market right now with their AI app or solution or whatever. But the ones that last are going to be the ones that are built securely, that are built performantly. And so I think if we did have this conversation in a year, we would start to look back and see, OK, like, yeah, that that project started off really hot, but it died because it just couldn't last because it ran into scalability problems, it ran into security problems. So we'll see who invests in that and who wins out because I think we're dispersed so early, we haven't even seen that pain yet.

And also when you don't architect your solution with the secure by defaults and with the performance pieces in mind, the tech debt is enormous. So they might not see that pain now, but they will later on.

James Cowling: Yeah, this is a great time for the security industry. This is this is the time the security industry has to step up and make security fun, make security appealing, make security desirable, and then we'll benefit a lot as an industry.

Allie Howe: Amazing. Well, super cool. Is anything on top of your mind for like what's next for Convex or anything else you want to say before we wrap up?

James Cowling: Every day there's something you were launching European residency in in 10 minutes. As soon as I get off this this call, actually what I'm going to say I'm not sure use Convex more than that. I need to hire people like we are growing right now. I desperately need more excellent engineers. And there's a there is a skill set that is dying out, sadly, which is the people who can design systems by hand, deeply understand abstractions and APIs this is a dying skill now. It's a skill. I think it's so important for the future of the tech industry. If this is you, if you're the kind of person who loves thinking about elegant systems design, come work here.

You'll have a really good time.

Allie Howe: Awesome, yes, Come work with James, it'd be awesome. Be fun, I'm sure was super excited to keep eyes on Convex and follow all of the happenings with Open Claw. And if you are using Open Claw, I'd love to know how are you building it securely? Maybe go check out Convex to work with Open Claw too. But yeah, we'll catch you next time. Thank you so much, James for coming on.

James Cowling: Thank you. Awesome.