Agents in the Database  ## Document Notes ## Highlights - **Sidelined Early Taught Management Lessons** - Ajay was sidelined at a company after an internal uproar and learned management visibility matters. - The experience taught him to both protect himself and to look out for his team as a manager. Transcript: Ajay Kulkarni I used to work at this company. And I won't mention you because I'm not trying to throw anyone under the bus. But I used to work at this company where I was in charge of an engineering team. And somehow, I thought I was on track to get more responsibility. And somehow, there was this internal uproar. And I realized people didn't like me. And it was uproar. And I ended up getting sidelined. And that hurt for a couple of reasons. Number one, I'm kind of a sensitive guy. Yeah, I'm an entrepreneur, but I'm like, look, dude, if you genuinely don't like me, I'd be like, why? Because I'm curious because I'm like, did I do anything to offend you? But also, I was just caught off guard. And what I realized is that part of that was my fault. Part of that was like, I think I came across more dictator-like than I intended. But part was, I would say, maybe indirectly my fault, where there were some decisions that were made where I took the heat for, where it wasn't actually, it wasn't the decision that I Made. I guess, I don't know, maybe, you know, I guess long story, I kind of realized that, you know, I think I changed my style. But I also realized that you have to look out for yourself, you know, and your manager, whoever, like they will not necessarily look out for you. But also conversely, I've tried to be the type of manager who looks out for people. ([Time 0:14:41](https://share.snipd.com/snip/6832cdda-38fa-4be9-8a47-0c4b95943e36)) - **Trust And Articulate Your Gut** - Trust your gut when hiring and making product decisions, then articulate it if needed. - If your instinct flags a mismatch, pause hires or strategies instead of rationalizing them away. Transcript: Ajay Kulkarni Made, you know, I'm sure you want to call them mistakes or just, you know, I've had learning experiences over the years. And I think that I've formed who I am today. I think the key theme, and this is not all of them, but I think one key theme is I have learned how to trust myself. Like, I remember early on, I would make some hires, example, in sales that didn't like gut feel right to me. But, you know, I was new to enterprise sales. I was new to the database business. And in hindsight, like my gut was right. My gut was right that like, hey, the general, you know, I like to say in a fast moving industry, expertise is a liability. So in the database world, as the databases were moving from enterprise sales and on-prem to cloud and PLG, the people who could sell this would be totally the wrong people for this. And I kind of felt that, but I couldn't articulate it. And I made a few hires that I had to let go. And that's my fault. That's on me. But I think that's one thing that taught me that, hey, you got this far for a reason, like listen to your gut. And if you can't articulate it, then try to take the time to articulate it. ([Time 0:16:50](https://share.snipd.com/snip/aeb8cc81-198b-4d93-9ea1-c7f85373b602)) - **Play To Founder Strengths** - Lean into your natural strengths instead of forcing balance or eliminating quirks. - Build teams that complement founder strengths rather than trying to change them. Transcript: Ajay Kulkarni But I think it's also a function of like don't try to shore up your weaknesses. Try to lean on your strengths. Yeah. Right? So like, I am an intense person. When I was younger, I would feel bad about it because some people are like, oh, well, you're intense. But now I'm like, no, it's just who I am. And some people love that about me. Yeah. There's some things I can do really well because I am in intense, you know? So I am, you know, like if you don't like that, but cool, you know, we don't have to work together. Right. You know, similarly, like, like I, you know, I talk a little fast. Like I remember at one point, like I was giving a presentation and like some of my peers were like, Oh, why don't you slow down? You're talking really fast. And I slowed down and they were like, no, this is worse. Like, it was better when you talk fast because your energy was there, you know? And I was like, you know what? Yeah, like I talk fast and maybe it's like, you know, maybe I might muddle some words sometimes. But like, that is who I am. That's how my energy comes through. And just lean on that. I haven't talked, I talked about it with Mike. It's like, we both have weaknesses. We both have flaws. And in the past, I would have been like, oh, we have to work on these. And now I've said, you know what? Like, don't try to teach a fish to climb a tree. You know, like if you're a fish, like swim better. You know, don't be like, oh, I can't climb a tree. It's like, no, just keep swimming, right? Just keep swimming. Yeah. ([Time 0:20:51](https://share.snipd.com/snip/d82b127e-2883-4798-a0c6-0b69cd474abd)) - **IoT Project Evolved Into A Database** - Ajay started Timescale as an IoT platform and built a custom database to solve his team's SQL pain. - A customer asked about the database, revealing demand and turning the product into a database company. Transcript: Ajay Kulkarni And it's like 2015. Yeah. I mean, back then, I just spent 10 years in mobile. And I remember thinking, wow, mobile was really exciting, but around 2014, started to get a little bit boring. Pre and post iPhone was exciting, but then 2014, you're like, okay, I have enough apps. I'm like, what's the next thing? And IoT felt like the next wave of computing. And so we started off building what we thought the market needed, which was a data platform for IoT devices. And that idea was moderately successful. We tracked over 100,000 devices. We raised a seat around, built a small team. But we needed a database to store all this data. And, you know, we were using a time series database. We were using a relational database. And I remember at one point, we wanted to, I think, sort the console by uptime, but then show all the device metadata. And what should have been a simple SQL join ended up being a two-week engineering sprint because you have to connect these two siloed systems. And I remember thinking, oh, this is awful. This sucks. And one of our engineers said, hey, I could build this on Postgres, but it'll take me a month. And I was like, okay, cool. You're an optimistic engineer, so it'll probably take you three months. And 12 months later, we had this database. And I'm trying to sell this IoT platform. And look, as an entrepreneur, you learn to listen to signals when someone is really engaged. And when someone's like, oh, yeah, that's cool. That's not cool. But when someone's like, wait, wait, wait, hold on. Can you tell me that again? You're like, okay, there's something there that's interesting to you. And so here I am trying to sell this IoT platform, meeting with the German head architect of this large shipping logistics company. I'm meeting them in Mountain View. And I'm telling him this IoT platform. And he was like, look, there's so many IoT platforms out there. What makes yours different? And I was like, oh, we built our own database. It does SQL skills for time series. And then he was like, wait, wait, hold on. Wait, can you tell me more about that database? And I think that's when I realized, wait, this is actually solving the right problem. And that's how it became a database company. So yeah, long story short, we kind of scratched our own itch and realized other people have the same itch. But again, we are students. Listen really well to the market. We try not to be dogmatic. ([Time 0:28:12](https://share.snipd.com/snip/26e230e7-5af3-4350-ab4c-511486e4896f)) - **Databases Must Serve Agents Too** - Agents change who the database serves: not just humans but humans plus agents acting on their behalf. - That shift requires databases to expose agent-friendly interfaces like CLIs and agent protocols. Transcript: Adam Stacoviak Yes. You know i think i'm at this point where there was this this uh this arc of acceptance i would describe it of hey come on now ai hype i'm over it too much this and that and i think the game changed When cloud code changed the game that's where i really think it happened like i had been a user of chat gpt like many people i'd been the api to do something in chat gpt and copy it out somewhere Else whether it was written material whether it was an idea whether it was a framework in terms of a thought framework or maybe even a bash script because ai is pretty good at bash scripting And then you kind of get to this other side where you're like you know what wow agentic is is really revolutionary and i think cloud could really change that game from the new browser or The new destination point for developers it is still in the IDE, right? And that's still taking place today. But I think you're seeing this shift to the CLI that is just truly revolutionary that now the terminal, it was never not cool, but it's cool again, right? A lot of folks are hanging out there. Cloud Code kind of put that on the map in a way. And then a lot of folks decided to follow that direction. And I imagine that's kind of what you thought, too, is like, I'm I'm probably doing you're probably doing a lot of agent encoding. You're probably playing with side projects. You're getting curious again. And you're like, well, the next best thing is to how can I just talk to my database? How can I just put an agent in my Postgres? Ajay Kulkarni Talk about that. Yeah, I think, yes. I had a very similar experience with Cloud Code. I remember one of my friends talking about agentic workloads in 2023. And I remember thinking, like, what are you talking about? What is agentic? What does that mean? And now I use the word agentic at least five times a day. Yeah. And CloudCode was that moment. CloudCode. I mean, ChatGPT was cool, but it was like, it was like a party trick. You know, it was kind of cool. Adam Stacoviak I could edit for me. For developers, it was a party trick. I mean, for everyone else, it's pretty powerful. But as a developer and ChatGPT, it's not the right interface. CloudCode. Yes, but not. Ajay Kulkarni I remember building. Because I was trying to talk to a friend. I was showing it to a friend. He was like, oh, can we build like an app that tracks pushups? And I'm like, I don't know. Let's try it. And like, I think 45 minutes later, we like chat. We Claude had gone out and found the right computer vision library and other stuff. And I had a mobile web app that would use computer vision to detect if you're doing a push-up or not. And it didn't get it always right, but it got it right at maybe 80% of the time, which was pretty good for 45 minutes of work. And I remember thinking, dude, I can build anything. Like I can build anything now. And I'm with you with like, I think Claude Code, it felt like, wow, this is actually an agent doing work for me, writing code, making decisions. I can steer it, making me more productive. I going to a social event that night and I couldn't even talk to people because I was so excited. Like I went home early. I was like, I just, I need to go back to Cloud Code. Adam Stacoviak I got to go. Ajay Kulkarni Well, I'm trying to talk to people about Cloud Code and they're like, what? Are you okay? I'm like, no, this thing's amazing. You have to try it. What kind of player was it? Adam Stacoviak Was it nerds or was it normal people? Ajay Kulkarni It was partners, part normal people. Okay. But even the nerds, but, but, but even the nerds were like, I mean, cloud code had just come out. Like, I think it was, there was still like, yeah, I use cloud. I'm like, no, not cloud, cloud code, you know, like, yeah, you know, no, I mean, like, you know, I told you about that experience in identity using the internet for the first time. Like, that's what this felt like. Like, number one, it felt like, I don't know what this is, but I want to be a part of this. Like, this is fun, right? But also, it brought me, and I know a lot of my peers that feel this way. It kind of brought out this like child, childish sense of wonder that I think like years of being of an entrepreneur had kind of beaten down and it kind of brought it back to the surface To be like, this is fun. This is fun the way technology should be fun where you're just like, I just want to tinker with this. I don't know what I can build. I just want to build things, explore this new world together. And yeah, so I'm really excited. I mean, I think as a company, we're excited because we see this trend that like, you know, 80% of Claude code is written by Claude. The majority of new software over the next 12 months will be written by AI. We already have customers who tell us that 70% of their code is written by agents. And you just look at this and you say, okay, if that's happening, then what else is happening? Well, that means the surface area of software development needs to evolve for agents, right? The surface area of databases. So now databases are serving a new user. They're not serving a human. They're serving a human using an agent. So then you ask yourself, okay, then how does a database need to evolve for that? Well, number one, it's less of a GUI. So you remember like databases went from on-prem to SaaS, like cloud, but now you're going from SaaS to, I don't know, MCP, CLI. It's like the interface is totally different. It's no longer like clicking right it's more like commands and there's some other things too like you want databases to boot instantly you want to be able to fork and create sandboxes Quickly in a safe way the cost-effective way you probably want native search native memory yeah i mean so i i don't know i'm just excited as as like the little kid in me who just loves loves Building things. He's excited, but also as a company, it's like, yeah, this feels like a problem someone should solve. ([Time 0:42:40](https://share.snipd.com/snip/e3496b4f-92d2-4d29-affa-f7cde6a68888)) - **Distribution Replaces Infrastructure As Bottleneck** - Lower build costs and AI tooling make shipping products dramatically faster today. - That shifts the main bottleneck from engineering to distribution and awareness. Transcript: Ajay Kulkarni It used to be hard to build, but now it's much easier to build, which probably means that the more of the bottleneck then becomes a distribution and awareness. Which is like, I think something similar to what, you know, the iPhone is. Adam Stacoviak It's always been a bottleneck though, distribution awareness. It's just exacerbated now. Ajay Kulkarni Yeah, but I remember once upon a time, pre-AWS, you probably remember this too, which is like, to start a company, you needed a server rack, you need maybe a couple of Sun, Micro, consistent Machines, Oracle, 8i database. And just to get started was such a high bar that a lot of people couldn't do it. I was actually going to say, I think it's similar to what the smartphone did to photography. Once upon a time, being a photographer was really hard, even if you look at before digital cameras, right? And you'd only take so many shots. You have to develop it. It would take time. It's hard to get right. You would have to actually apply physical filters and physical lenses to change the thing. And with the smartphone, everyone became a photographer. And what that meant was that in the past, having great photos would set you apart. But now it's like anyone can create a great photo. But how many followers do you have on Instagram? So it kind of again, shifted the burden from photography skill to like distribution. And I feel like we're going to go through something similar with software. ([Time 0:54:37](https://share.snipd.com/snip/b04de2bf-7c53-436f-8cf3-e6e2e17218cf)) - **Design Tools As Agent Affordances** - MCP and CLIs act as affordances for agents, teaching them how to use services. - Good tool design for agents requires deciding power, sandboxing, and concise capabilities. Transcript: Ajay Kulkarni I know the team that built it is really smart. You know, I know, I think they looked at a couple of different options. I think the way it works right now is that the CLI will run a local MCP server, which then speaks to a remote one that we have. And the remote one is where we index, you know, our MCP will allow you to do things with Postgres and Tiger, but also lets you, you know, search the Postgres documentation, download best Practices. And so we put that on its own kind of remote MCP server. I think MCP in general is, it's such an interesting area. And I feel like a lot of people hate on it because, seems so simplistic. But I think just the idea that people talk about in the physical world, the idea of a design affordance, this concept, it's this idea that when you see something, when you see this mug And you see this thing, you know you can hold this. It's kind of like a door handle. You're like, I know I can turn this. There's one reason why Tesla door handles trip some people off because they don't see the handle. What's going on here? Yeah. I think that's bad design. But in affordance is when the tool teaches you how to use it because it's designed the right way. Adam Stacoviak Like games, like Mario. The Super Mario Brothers, the first game, that was why it was level one was level one. It was teaching how to play level two and three and four. Ajay Kulkarni And so this is where MCP to me is not just an API. It's essentially a way to expose tools for agents. To essentially think about like, hey, we used to make these handles for human hands, but now we're going to make them for agent hands. And that is a really growing art. ([Time 0:57:24](https://share.snipd.com/snip/ecb2563d-6c58-4867-93ca-4b35d14ed448)) - **Give Agents Contextual Docs And Commands** - Provide agents with context by indexing docs and best practices alongside CLI tooling. - Combine CLI commands and semantic docs so agents act like expert developers out of the box. Transcript: Ajay Kulkarni I think the key thing is the CLI is a clear program that agents can use. So, okay. So when we built up Tiger CLI, like it's written in Go, it's distributed as a single binary. It essentially provides a unified control plane for your database on Tiger Cloud. And so, you know, it covers, you know, auth obviously creating, starting, stopping, forking databases, connecting databases. But like what's really, think, nice about it, kind of like you pointed out, is that it has MCP baked in. So, you know, you can run essentially a local MCP server through the CLI that you're then your cloud code or whatever can talk to. And now your agent has access to all these basic commands. It has access to starting, stopping, forking databases, connecting the databases. It also proxies another MCP server that we host where we have indexed the Postgres and timescale and tiger documentation. And now we provide semantic search and tech search on there. So again, it's not just a CLI. It's not just a CLI MCP server. It's a CLI MCP server that actually makes your agent smarter. I don't know. I would say, I'm not sure if this is true, but it probably is. ([Time 1:07:25](https://share.snipd.com/snip/965dcfe3-40da-4427-818e-51070bafd69c))