Tencent researchers built and tested NarrativeSteward, a system for interactive fiction authoring that divides work between human writers and a team of specialized agents. The core finding is that writers need three distinct controls. They must be able to hand a task to an agent with a clear goal, watch the agent work in real time, and reclaim authorship of any piece the agent got wrong. Without all three, writers either hand over too much and lose their voice, or refuse to hand over anything. The system uses a structured layer in which the writer sets goals and constraints before the agent starts, then reviews a plan before any content is generated. Agents signal their confidence and surface their reasoning so the writer can catch problems early. A dedicated recovery path lets the writer roll back any agent decision to a prior state and resume from that point. The agent is aware of the correction when work continues. For product teams, the key takeaway is that handing work off and taking it back are not separate features. A writer hands work to the AI only after believing recovery is possible. NarrativeSteward makes that belief credible by keeping each agent decision visible, reversible, and attributed. The paper includes a user study and a public release of the code.