An operator set his AI agent an exam. It failed
The order matters: teach first, test second, trust last. One founder writing for Inc. put his most important agent through a written exam before letting it near real work.
The practice is borrowed from hiring. Before an agent is trusted with work that matters, it sits a test.
The founder, writing for Inc., describes the sequence as teaching the agent, then testing it, then trusting it. The agent failed the exam.
That is the useful part. A failure on a written test costs nothing; the same failure inside a customer workflow costs a customer.
The habit is cheap to copy. Write down the tasks an agent will actually be asked to do, set them as questions with answers you already know, and mark the results before deployment rather than after.
Most teams skip straight to trust because the demo looked convincing. A demo is not an exam. It is a sample chosen by the seller.
The pass bar is the founder's to set, and setting it in advance is the discipline. An agent that has not been marked against a standard has not been tested at all.