Most agent projects start with a convincing demo. The agent reads a ticket, looks up an order and drafts a reply. The gap between that demo and a system you can leave running overnight is where most of the work sits.
The first step is evaluation. Before changing prompts or models, we collect a few hundred real cases from the client and score every version of the agent against them. Without this, every change is a guess.
The second is tool design. Agents fail most often when tools are vague. Narrow tools with clear inputs, such as "refund order up to $50", are easier for the model to use correctly and easier for people to audit.
The third is handoff. Every agent we ship knows when to stop and pass the case to a person, with a summary of what it already tried. That is what makes teams comfortable letting it run.