A go-kart has an engine, wheels and steering. That is genuinely a vehicle. It is also not allowed on a public road, and nobody needs that explained to them.

An AI agent is three things: a model, a set of tools it can call, and a loop that runs until the job is done or the turns run out. You can build one in an afternoon, and the one sitting in your prototype repo right now is probably exactly that. It works. It demos well. Somebody says the word transformative and they are not being silly.

Then it gets pointed at production, and nothing about it changes except what it is aimed at. Real customers, real records, and a service account that opens every door, because that was the quickest way to make staging work three months ago.

The failure, when it arrives, is not a bug. The agent calls a tool the user could never have called, against a record they could never have opened, and it does so because the permission said yes. No hallucination, no jailbreak. The postmortem runs to four sentences and none of them are about the model.

The infographic maps twelve parts of a road-legal car onto twelve parts of a production agent, and follows one go-kart through the seven stops between a garage and an open road.