Thinking vs. Doing: Agents that Reason by Scaling Test-Time Interaction