Replit's Michele Catasta releases ViBench to measure how effectively AI agents build and edit complete applications

It uses anonymized Replit applications converted into product requirements