For agents, the deciding capability is rarely raw knowledge; it is whether the model reliably respects tool schemas, stop conditions and formatting contracts across dozens of turns. A model that drifts from its contract at turn fifteen cannot be deployed unsupervised.