Closing the loop: Evaluating and improving Replit Agent at scale

Most Replit Agent users start with an idea. They describe the goal in natural language — without a repo, test suite, or chosen framework — and expect the agent to turn it into a functioning app.