Inside the Y Combinator Browser Use Agents Hackathon with Convex
It's Mike from Convex and we are at the YC web agents hackathon. It's going to be one of the craziest hackathons. All these applications are going to be fire. Why? Cuz they're using Convex. There's a ton of other great sponsors, but none of them are better than Convix. Let's get inside. Convex. Yes. Say that one more time. A lot of excited people trying to get some credits, trying to win some money, potentially trying to get a job. Anyone heard of Convex? Convex. Okay. We're going to make the greatest convex application ever created. The greatest convex application is able to take any influencers's handle on Tik Tok. Custom build them their very own business product. Business store you can make some money on. Oh yeah. Claude or Codex if you have to pick. Claude is your adventurer. CEX is your reliable companion. He's the one that's actually going to build the infra architecture. Make sure it's reliable and elegant. So you need both. Oh yeah, we need both. I like Wait, it's 100? Yeah. 160k $160,000 could be in someone's pocket. Wait, I can't I can't take part. We can't take part. You're easy convex, my man. [music] You see, I like What up, dude? My dog. We were looking through the internet and just like there was this post we saw where a pregnant woman was like, "I'd rather go through 36 hours of labor than move." Honestly, it is a painful thing. I helped a friend move recently and it was the worst thing ever. Perfect example of just like where agents can automate this whole process. Our whole thing is you can use OpenClaw and Convex DB where OpenClaw writes to Convex, stores all the context for your move and just coordinates the whole thing across all these different services that you need to do. Convex is going to help us just because like we need to be able to store the context of [music] all these different like small interactions between all the agents that we have and everything is going to be real time too cuz a lot of bozos are not using it. So y'all got that advantage. Are you ready? Are you all right? [cheering] Like this is an opportunity to change your lives. [music] Where do they write the checks? That's where I want to be at. Write the check. You know, like how many deals have been made through this little phone? Act like I'm getting in. We were talking about being performative earlier. Oh, my time's now. My time is now. Yo, Dan, can you hear me? Oh, yeah, you can, cuz I got the Oh, it's really good. Dang. I need a raise. Oh, wait. I might have said something that should not be on camera. [music] I'm hungry. Oh, they got salads, bro. Yeah, we getting food outside, bro. They got kale salad. I look like I eat kale bro. You're walking up there. Hold on. Let me get one more bite. Give me one more. That sounded good, my boy. Dang. I'm sure you recorded me hating on it earlier. See, we were doing a YC hackathon. Me and one of my good friends who's also here [music] today. And we ended up winning first place overall at YC in August. After that, we met the browser guys and it's kind of just been smooth sailing from there. And we also have some like the coolest startups in this current YC batch as well, just showing up and making something really awesome. Like the way you guys set up like V and a front end with the Convex API and database already set up for you is amazing. And that kind of makes it so that you can spin up a hackathon project or a side project in literally like 5 seconds. The way I got [music] my convex job, I want a hackathon. I use convex fully. Yeah. And they like the project. So maybe if you're looking for a job or hackathon is the way to do it in your email where like your addresses are stored. So like all these different sites and then then it'll start searching for like where to live. So like based off your budget and stuff, it'll go find some places. Tell it that you don't have a place to move. Your budget's $0. It'll tell you the best corner to live on cuz like Yeah. No, this is this is an app built off pain. Oh, dude. It's all like three prompts. So, yeah. Yeah. You got a prompt. Got a prompt, Max. All you need is that cloud coach skill. You just use whisper and then you just keep talking and talking. So, you guys are not doing the boomer way of writing code by hand? I I don't even code. [laughter] There you have it. I'm a civil engineer. There you have it, legend. Well, y'all keep cooking. What do you think? I don't know what's I I have no clue what's happening. What do What do What do What do What do What do What do What do What do What do What you think is hell doing B- roll? It's like that. Yeah. Like I think he's posing for his new album, right? He post for his new album. Every Sunday I go to church. There was not a single Sunday after church. Did not meet anybody working on AI agents. You always wish before you go to a meeting, of course, you want to your research, you want to look up who is the person, but you never do it because of time. You just prompt this A2 every day 6:00 a.m. ping me on Slack my schedule for [music] a context. I was building this like the best nei system in the world and now the agents world has evolved cuz now it's like perfect time. But we're probably going to go the convex route and actually put out something open source that people can truly own their data and everything. Essentially an automated but for browser for example specifically a big problem with a lot of legacy websites is the lack of standardization. So this is the lobby here on the left hand side. You can pick your fighter bridge. I think it's trying to purchase me some toothpaste for a rooting brain for web agents. Think of it like waves but for AI navigation on web game super hard. And then I just thought what if I had a loom which could actually execute it for her. So that's [music] what this is. Now I want to show you especially in Europe the um energy market is very fragmented. It lets you [music] detect anomalies, track your agent behavior, and get notified before critical events become large scale incidents. We're going to show you a quick demo. First place, give me the biggest drama out of this. You guys do browser brawl. How do you both feel, dude? Fantastic. I can't even describe it. Their face was like uh uh and then it was like where did the idea come from? First and foremost, we were both really interested in like the data spot. I spent the last couple of days thinking a little bit about what's the next frontier for agents. I really like this idea from earlier Google like agents playing against agents and then you know like papers like the Jan paper generative which is how old is that paper? It's 2014 paper. Yeah. Yeah. So so he's talking about some random 2014 papers like yeah sure I guess I thought we could make we could apply them to the browser. It would be really sick. He pitched me like this game. At first I'm like, dude, lock it. Like we're trying to win this. But then I started thinking about it and I'm like it might be crazy enough to win this. What was that experience [music] like using and building with comics? I think the name of the game with these hackathons with speed over everything. And I'm an AWS guy formerly, right, with these databases, maybe super base, but it was just so fast like unbelievably fast. It just works with cloud code. I was able to spin it up. We liked it so much we ended up blowing through that free tier. We end up asking you like, "Yo, can we get some more?" And you guys came through. You guys came through. We pitched the idea to Lucas from Anthropic. He was like, "Yeah, this idea sounds really good, but I think it's a research rabbit hole. You should definitely not do it." And then we went and definitely did it. [laughter] Young guys, bright [music] future. Shout out to you guys. I'm excited to see where you guys take it from here. Thank you. Whenever you think, oh, other people are building this or Yeah, for sure. Open eye or in topic, they will they will figure this things out. So, it's not really worth it to to build this. what's actually in your heart. You should you should just think how hard could it be and just stop it.
A woman on the internet once said she'd rather go through 36 hours of labor than move apartments again. One team at the Y Combinator Browser Use Web Agents Hackathon built their whole weekend around that line. It's as good a starting point as any for what happened in San Francisco: a room full of developers pointing AI agents at the parts of life that are technically solvable but nobody wants to do.
We sent our DX engineer Micky (RasMic) to the hackathon, hosted by the Browser Use team (founders Gregor Žunič and Magnus Müller, with founding engineer Reagan Hsu). The brief was simple. Hand people Convex, Claude, Codex, and a browser-automation library, then see what they build. What came back wasn't a tidy list of demos. It was a room full of people who'd all landed on the same infrastructure decision, independently, for the same reason. They didn't have time to think about their backend, so they picked the one that meant they didn't have to.
The moving problem
The team building around that "36 hours of labor" line had a real, specific complaint. Moving is miserable, and most of the misery is coordination overhead: contacting movers, chasing quotes, tracking who said what. Their pitch was to let AI agents do that coordination instead of a person.
The architecture was OpenClaw writing to a Convex database. That database held the contacts and interactions across every small exchange the agents had to track: movers, listings, logistics, all of it. As one team member put it, "we need to be able to store the contacts of all these different small interactions between all the agents that we have." The real-time piece mattered too. Most competing teams weren't building with anything reactive, and that's the kind of default advantage a synchronized backend hands you when you're not the one building the sync layer by hand.
Winning by not thinking about the backend
The line that came up more than any other, from multiple teams independently, was about setup time. One repeat YC hackathon winner (his team took first place overall at a previous YC hackathon in August) put it plainly: "the way you guys set up a front end with the Convex API and database already set up for you is amazing, and that makes it so that you can spin up a hackathon project or a side project in literally five seconds."
That's not a small thing at a 24-hour hackathon. The honest allocation of your time there is closer to "twenty hours of building and four of arguing about whether Claude or Codex should own the harder parts of the stack" than to any leisurely architecture review. One developer's framing of that debate stuck: "Claude is your adventurer. Codex is your reliable companion. He's the one that's actually going to build the infra architecture, make sure it's reliable and elegant." Their team used both. Claude for the unknown parts of the problem, Codex for making the known parts solid. That's a more specific answer than "we used AI," and probably a better one than picking a single tool.
Another developer went further. The win itself, he said, came out of a hackathon loop: he'd won first place at an earlier YC hackathon using Convex, and the project was good enough that it turned into his job. "I got my Convex job because I won a hackathon. I used Convex fully. And they liked the project, so maybe if you're looking for a job, a hackathon's the way to do it."
Finding a place to live for free
A second team built a tool that stores your email addresses across sites, then searches for housing based on your actual budget. The demo line was blunt: tell it you don't have a place to move and your budget is zero dollars, and it will still tell you the best corner to live on. It's a joke. It's also a real demonstration of what agentic search is supposed to be good at: working an underspecified, unpleasant constraint problem a person would otherwise do by hand across a dozen open tabs.
What's worth noting about that team's build isn't the app, it's how little of it was hand-written. "Three prompts," one of them said. Their teammate, self-described as "not even a coder" and a civil engineer by training, had a working app anyway. The team's workflow skipped typing code in favor of talking to a coding agent through Whisper and letting it write. Whether or not that's the future of software, it's clearly workable enough to win a demo slot at a YC hackathon in 2026.
Automated meeting prep, and a naming problem
Not every idea at the hackathon was about moving or housing. One recurring pattern showed up across several unrelated conversations: using an AI agent to do the unglamorous research you'd do before a meeting anyway, if you had time. "Before a meeting you want to research who the person is, but you never do it because of time, so you prompt an AI: every day at 6am ping me on Slack my schedule with server context."
There was also a smaller, funnier thread running through the day. Several people separately observed that nobody has settled on a name for this category of software yet: agents, agentic apps, web agents, take your pick. Not knowing what to call the thing they were building didn't stop nine teams from demoing it in front of a judging panel by evening.
The winning demo
The first-place team came from a place most hackathon origin stories don't. Not "we saw a market gap," but "we were both really interested in the data space" and a few days spent asking what the next frontier for agents actually was. The idea that stuck: agents playing against agents. It drew loosely on the generative agents research going back over a decade, applied to the browser instead of a simulated town. "At first I thought we're just trying to win this, but then I started thinking it might be crazy enough to win," one of the builders said. That's a more honest description of how good hackathon ideas arrive than most retrospectives admit.
Their account of using Convex under time pressure was specific rather than promotional. One team member, previously an AWS user who'd also worked with Supabase, called Convex "unbelievably fast" to get running. He credited it working smoothly with Claude Code for speeding up the build, and said the team liked it enough that they blew through the free tier and had to ask for more. That's a real constraint hitting a real product decision, not a talking point.
The team had pitched the idea to Lucas from Anthropic beforehand, who told them it sounded good but was "a research rabbit hole" and that they probably shouldn't attempt it in the time they had. They built it anyway, and it won.
What the room actually agreed on
Strip away the specific apps and there's a pattern underneath all of them. Nobody at this vibe-coding hackathon spent their limited hours deciding how to store data, wire up real-time updates, or provision a database. They spent it on the part of the problem that was actually new: coordinating agents, parsing browser state, deciding what a web agent should be allowed to do on your behalf. The backend question was answered before the clock started. That's the real argument for why a database that gets out of the way matters more at a hackathon than almost anywhere else: the honest cost of infrastructure isn't what it does wrong, it's the hours it takes from you even when it works.
As one participant summed up the mood of the whole event, only half-joking: "We used to think other people would build this, OpenAI or Anthropic would figure it out, so it wasn't worth building. But what's actually in your heart? You should just think, how hard could it be, and start."
Build in minutes, scale forever.
Convex is the backend platform with everything you need to build your full-stack AI project. Cloud functions, a database, file storage, scheduling, workflow, vector search, and realtime updates fit together seamlessly.
