What AI agents actually need before they touch production.
An agent demo and an agent in production are two different products. The gap between them is almost entirely about failure handling.
Every agent demo looks the same: a clean prompt, a clean tool call, a clean answer. Production traffic is not clean. Users paste in half a question, the API they depend on times out, the data has a typo in the one field the agent needed. The agent that looked impressive in the demo has to survive all three of those on the same afternoon, and most were never tested against any of them.
The fix is not a smarter model. It is a narrower one with better guardrails. We scope every production agent to a small, explicit set of actions it is allowed to take, give it a clear way to say 'I don't have enough information' instead of guessing, and put a human checkpoint on anything that touches money, customer data, or an irreversible action. That combination is less impressive in a sales deck and far more reliable in week six of actual use.
The teams getting real value from agents right now are not the ones with the most ambitious scope. They are the ones who shipped a boring, well-instrumented agent for one workflow, watched what it got wrong for a month, and only then expanded it. Evaluation harnesses and logging are not optional infrastructure for agentic systems — they are the only way to know whether the thing is actually working once it is out of your hands.