Agents are loops

An agent is not a personality. It is a loop: model → tool calls → observations → model, until a stop condition. You own the stop condition. Max N steps. Max $Y. Max wall clock. If it cannot finish, it should return a partial with a reason, not spin.