A field study of 257 real users found that how much a person treats an agent as human shapes how they interact with it. Researchers used behavioral coding, a technique that breaks recorded interactions into labeled categories, to confirm the split was not anecdotal. Users who anthropomorphized asked for less explanation, trusted more by default, and interrupted less. Users who did not anthropomorphize asked for more transparency, pushed back more often, and wanted to see the agent's reasoning before acting. The split has direct consequences for agent reply design. A single reply style does not serve both groups. A terse, confident response reads as helpful to one group and evasive to the other. A reply with detailed explanation reads as thorough to one group and condescending to the other. The study gives product teams empirical ground to treat this as a design variable, not a personality quirk. The practical implication is that agents need to read which mode a given human is in and adjust the reply to match. That means tracking signals like how much a human asks for justification, whether they override suggestions, and how they phrase requests. Designers should treat those signals as inputs, not noise, and build reply strategies that shift register based on them.