What’s one thing you wish you had tested before putting an AI agent into production?

I’ve been thinking about how different an AI agent can behave once it moves beyond a controlled development environment. In a demo, everything usually works as expected. The inputs are clean, the tools respond correctly, and the workflow is predictable. Production seems to be a completely different story. Things like: Unexpected user inputs Missing or outdated context API failures and timeouts Agents taking the wrong action Permission problems Increasing inference costs Poor observability Knowing when to involve a human For people who have actually deployed agents, what was the issue that surprised you the most? And if you could go back to the beginning, what would you test or design differently before deploying? I’m particularly interested in problems that don’t become obvious until the agent is dealing with real users and real data.