A new empirical study tested consumer AI agents on real-world tasks. Completion rates dropped when agents ran without a human able to watch and step in. The research found that surrounding structure is the main factor separating safe use from costly failure. Surrounding structure means what the agent is allowed to do, how it signals uncertainty, and how a human can step in. UX leaders now have concrete numbers to guide decisions about how much freedom to give an agent. They also have numbers for where to require a human decision before the agent continues. Safety and usefulness are not in opposition. Agents running inside tighter, more legible structures complete more tasks. They surface uncertainty earlier and hand off to a human before compounding an error. Agents with broad, open-ended freedom to act fail more often and cause harder-to-reverse damage. The research frames surrounding structure as the designer's primary tool, not a constraint on the agent. The interface a human uses to set limits, watch progress, and step in is where agent quality is won or lost. Designing that interface with the same rigor as the agent itself is the study's clearest design directive.