The field has moved three times in three years. First we argued about how to phrase instructions. Then we argued about what to put in the context window. Both mattered, and both stopped being enough the moment people started shipping agents that act on their own. The formula everyone converged on is simple: an agent is a model plus a harness. The model supplies reasoning. The harness supplies everything else, and the harness is what you are actually paid to build.
This course teaches you to build one.
What You Will Learn
We build a single system end to end, a Support Resolution Agent that reads a real customer ticket, looks up the order, decides on an action, and answers. You start from a bare API call and add each layer of the harness in turn: the agent loop, the tool interface, context delivery and compaction, memory and state, permissions and guardrails, verification, tracing, and a human approval gate for the actions that carry risk.
Along the way you will meet the failure modes that only appear in production. Agents that declare victory on work they never verified. Agents that get sloppy and rush as the context window fills. Agents that try to one shot a problem that needed four careful steps. None of these are prompt problems. Every one of them is a harness problem, and each has a concrete fix you will implement yourself.
