A study of 30,000 AI agents and interviews with the humans who built them finds that variety collapses over time. The cause sits in how humans interact with agents, not in the models themselves. Humans settle into a narrow set of prompts. Agents then produce narrower and narrower outputs. The researchers call this semantic collapse, meaning the range of what an agent does shrinks until it feels repetitive. The study identifies two forces that slow this collapse. First, humans who vary their instructions, add information before the agent starts, or shape what the agent knows keep getting varied results longer. Second, interface design either helps or hurries the collapse. Products that make it easy to repeat a single prompt, without showing what the agent already knows or has already done, push humans toward repetition. Products that show what the agent is drawing on give humans a reason to adjust. For a team building an agent product, the lesson is that variety is a shared responsibility between the human and the design. A product that shows what an agent knows, where that knowledge came from, and what it has already tried gives the human something to act on. A product that hides all of that leaves the human with no foothold, and the agent drifts toward sameness.