Another thing we do here to drive quality is avoid context rot. Agents get dumber as their context window gets bigger. To combat this, we explicitly ask the agent to do the smallest unit of work in our implementation prompt When done we will run that work through lots of review and CI steps before than in a fresh context window doing the next thing

One of the major themes of how I architected the agent building this is the agent spends way more time refactoring testing and benchmarking the existing code than it does adding new fetures. Only ~1/10 context windows adds a new feature

Now proudly introducing, the first ever vibecoded Ethereum execution client It's actually 2 clients - Zig client - @EffectTS_ client If you want to see how the AI Slop Sausage is made, follow XVI 1/2