The test is the executable specification

Writing tests before code has always been good practice. With agents, it stops being optional, because the test is the only way for the agent to know whether it got it right without asking a human.

Without a test, every step requires someone to say, "that looks good." That does not scale, and the agent starts consuming the attention of the person who should be doing something else.

The test becomes the target: the agent runs, sees red, fixes it, sees green. Ambiguity also shows up in the test, because a criterion that cannot be written as a test is usually a criterion no one had actually defined.