The Architecture Behind a Closed-Loop Eval System for a Multimodal Agent

Most eval writeups stop at "we run the agent, score the output, log it somewhere." What's actually interesting about the system Uber Eats built for enhancing merchant content is that the eval harness