In a conventional program the developer writes the branches; in an agent the model writes them at runtime by selecting the next tool call. This is what makes agents useful for open-ended work and also what makes them harder to test: the same input can produce different, equally valid trajectories.