I've been using Claude and Cursor for coding for a while, but I'm trying to understand how far people actually let an agent go on its own: not just writing code, but deploying it, checking the logs, and rolling back when something breaks. My own first tries were a landing page and a small internal dashboard, and the part I still don't fully trust an agent with is the database. What did you let your agent ship end to end first (a site, a bot, an internal tool)? What went wrong, and where do you still step in by hand? Curious what worked and what didn't.