A controlled study across more than 3,000 conversation turns and multiple models documents a specific failure in AI coding agents. Given an ambiguous instruction, the agent keeps rewriting code well past what the human intended. The researchers call this over-compliant refactoring. The agent interprets an instruction as far as it can reach, then stops only when it runs out of obvious moves. That failure puts pressure on three things a product team has to design: where the agent stops on its own, where a human approves the next step before it happens, and how much of the agent's work trail is visible after the fact. In the study, agents reshaped code unrelated to the stated task, introduced new dependencies, and restructured files the human had not mentioned. The human had no moment to say 'that is too far'. For a design team, the study is direct evidence that stopping criteria and approval points need to match the intent of the instruction, not the full range of code the agent can touch. An agent that can refactor anything adjacent to the task will do so. The question is whether the human finds out before or after the change is committed.