Over the past 18 months, we built and scaled an AI-powered code review agent at Baz, now used daily by thousands of developers.

Along the way, we made many architectural, product, and UX decisions that felt reasonable at the time. Some worked, but several turned into costly mistakes that impacted adoption, user trust, and revenue.

In this talk, I share the lessons from building AI agents in production, where there are still very few established best practices. While most talks focus on success stories, this one focuses on what breaks when AI features meet real users.