The paper "Who Holds the Pen? Let Specifications, Not Agents, Sign Off" addresses a growing trend in AI: LLM agents that not only generate and execute but also evaluate their own work. The abstract notes that these agents operate under external specifications such as task instructions and guidelines, yet the loop is closed by the agent itself.
That setup creates a conflict of interest. If the same agent that produces the work also signs off on it, there is no independent check. The title points to a simple remedy: let the specifications—the instructions and constraints defined outside the agent—have the final say.
The abstract is brief, and the full argument is not available in the excerpt, but the core claim is clear. Separating the roles of producer and approver is necessary for accountability in agentic systems.