An agent is a model called repeatedly, with each tool result added to a growing transcript until something ends the run. That replaces “how autonomous is it?” with three questions that have answers: what ends the loop, what one turn costs and what one turn can touch.