# Vibehouse: The New Ethereum Consensus Vibecoded Client | Lion Dapplion - Lighthouse

- Channel: [Ethereum Denver](https://streameth.org/ethereum-denver)
- Date: 2026-03-09
- Duration: 09:43
- Topics: ETHDenver, Crypto, Web3, Blockchain, Event, Conference, ETHDenver 2025, ETHDenver 2024, Bitcoin, Ethereum
- Watch: https://streameth.org/watch/yt-9XRQR3HsmO8
- YouTube: https://www.youtube.com/watch?v=9XRQR3HsmO8

## Description

Vibehouse is an Ethereum consensus client built almost entirely by AI agents. Forked from Lighthouse, it shipped a complete ePBS implementation in under 72 hours — all spec tests passing, finality on a multi-node devnet. This talk covers how vibecoding works for consensus-critical code and what happens when you let AI ship a hard fork.

## Transcript

Hello. All right. Our next talk is slightly different from what's on the agenda as it changed last night. Um, it's going to cover how WB coding works for consensus critical code and what happens when you let AI ship a hard fork. And please join me in welcoming Dap Lion to the stage. Hello, I am Dub Lion Cord at Lighthouse and uh yeah, I joined the talk last minute because AI is going crazy and we're running a super cool experiment that I think was fit to show here. So I'm officially presenting Bhouse which is a new consensus client fork from Lighthouse and fully written by AI after the previous fork. So Lighthouse is completely agentic generated from the previous tables version which is uh 8.0. It's now very simple. Just have a simple Ralph loop trying to implement the fork, pass spec tests, run defines, and it's also open to contributions. If you open an issue or a PR, it will do its best to consider and implement it. And we don't have a road map, we have a buy map. So it development speed has been insanely fast and it's been able to successfully implement Glowas in a matter of a day or two. [snorts] and also interop with another bipod code client. So b house as of today it's glow as ready. It passes all the spec tests. It passes curtosis. And as of today, it's already doing some uh bests like it's implementing a feature that was requested by the community. Super interesting, which is incorporating ZK execution proofs to make clients stateless. And it's doing this on the background as we speak. So this is the fun part. Doing bite house has been super fun to run this gambling machine and see what happens. And so far it's been massively successful, but for real, what the [ __ ] is this going on? Why are we doing this? So in Lighthouse case, our our bottleneck are reviews. We have this one or two uh very very senior engineers that need to review everything and that's what's limiting our bandwidth. As you know, Ethereum development speed is quite slow. We are shipping about one fork every year and we believe we can do better. Even inside of Lighthouse, we feel some frustration by the development speed and we are really optimistic and how AI can improve uh this development speed like now we have uh even a collaboration with the testing team to see how we can speed up development. So what Bhouse is giving us is we have a commitment in Sigma Prime and Lighthouse to always ship the secure code because Lighthouse has never had a bug in production in all its lifetime since Genesis and we will keep doing that. There is always a human in the loop. So Vive House is this parallel playground of just go crazy just do whatever you want with AI and then we want to see where this thing breaks. That's going to give us a hint where we need to increase testing efforts and where AI fails so we can improve that. Also, something I'm personally very excited about is to break this gatekeep model that we have with clients. Like the process to improve Ethereum, it's very slow. You need to do an EIP, get consensus, then the clients will consider and then they will implement it. This can take weeks, months, and even something simpler like adding a new API call to the clients takes a long time. What would the look if you are a nunner, you need something, you just go to the client, you open an issue and it gets implemented in minutes. How [ __ ] awesome can that be? And I think we can get that with like proper agentic development. So I think that this can like I don't know create a new golden era of relationship between [clears throat] node runners and clients and then all of these lessons we can port them back to Lighthouse safely and improve our processes. So what consensus clients have that's really good for them is we have an extremely robust spec along with spec tests that have an insane amount of coverage. And this is the guiding candle, the guiding carrot for these agents to successfully implement these experiments. We also have an amazing amount of tooling curtosis and other and other ways to run defets that make the process seamless. So this is what these agents has been leveraging and they have succeeded and now even the testing team is incorporating AI to do even more tooling that's even more extensive doing even more spec tests. So and that's here is where we can see this scape velocity where agents can improve exponentially. So in Lighthouse today our adoption of AI is I would say rather small. Most of the code that we have is written by hand. There are maybe like two or three teammates that are using clot as an assistant to write a little bit of code. But we can go much further. What we are doing now is we're starting to use AI to write tests to increase testing and especially to to test this these features with devets. AI is massively massively useful and I'll show you now. So the vision with this is what if instead of shipping one fork a year we could ship three. We could ship four. What if the testing cycle instead of taking one definite every two weeks could be a matter of days or even hours. I think we can get that if we somehow can increase testing coverage and even like the speed that we can fix bugs. we can get there. Okay, I'm missing some slides. So, I'll just talk over this about what we're going to see. So, I was going to show a practical example of how this uh testing loop can improve. So when we developed Ethereum, how it works is someone writes a a spec, we produce spec tests, then someone implements a spec and then we schedule a defnet which could happen maybe a week or two weeks later. Then we run the defet falls, then the Barnavas or Par or someone from the EF goes to Discord and be like, "Yo, uh, Dab Lion, we have a bug. Can you please fix it?" But I'm asleep. So I wake up 24 hours later or somewhere. I fix the bug, push a branch, then they run the devet again, but they need someone from guest to fix it, which maybe they don't wake up in 12 hours and then they fix the bug. And then we schedule the devet for like one week later. So the deathnet cycle is quite slow and it it it's also tied to the to the calls that we need to coordinate all of this if we need some synchronicity. But with agents, I'm not developing the code. My agent can do it. And what we're doing in Lighthouse is that we are committing all the instructions, all the tests, all the files that we can so that anyone can write code, not only us. So now when the testing team runs a DevNet and they find a bug, they don't need to wait for us to fix it. They can just fix it themselves and they can open a PR to Lighthouse with a fix. They can test it and they can even merge it in a fork. So really the speed of development now would be they implement the spec an agent writes the spec they run the devet if there is a bug they fix it and they then they can PR back to Lighthouse. So what used to take three weeks now it could be just a matter of of hours and that just like [ __ ] blows my mind and we can see this maybe in the next coming weeks. This is not something farfetched. This is something we can do today and this would massively massively improve the speed at wi at which we can develop. It's also scary though because now it means that anyone can develop clients and one of the scary parts of it is that if you have an idea for a buck or a malicious client, you can just bite code it and have a malicious client messing with the network. Um, so I hope that AI doesn't give us more work than we have now. Uh but yeah, it's interesting arms race that we will have to deal with. So yeah, Bhouse is this amazing experiment that we're running now to test how far we can push this technology. And I would encourage you or anyone watching, if you have an idea for Lighthouse, if you ever wanted to have specific flag or specific feature or something, just go open an issue and participate with this experiment. Thank you.