Preserve, clarify, infer, decide, ignore: an information-responsibility boundary for human-directed agents

Human-directed agents increasingly execute open-ended tasks from natural-language instructions, yet those instructions rarely specify every task-relevant choice. During execution, an agent must repeatedly determine whether to preserve stated information, ask the human, infer from context, decide autonomously, or disregard non-operative context. Outcome metrics reveal whether a task succeeded, while descriptive trajectory taxonomies reveal what actions occurred; neither alone determines whether the agent assumed the right informational responsibility relative to human intent. We introduce the Information-Responsibility Boundary, a conceptual model of this allocation, and five responsibilities: $\textit{Preserve}$, $\textit{Clarify}$, $\textit{Infer}$, $\textit{Decide}$, and $\textit{Ignore}$. The model originated in product-oriented text-to-app development but applies more generally to human-directed agents operating on partially specified intent. We position it as a semantic layer complementary to Act$\cdot$ONOMY's descriptive taxonomy of observable runtime behavior. We further show how responsibilities change as interaction updates the information state, enabling trajectories to be interpreted in terms of responsibility transitions and violations. We conclude with an agenda for trace coding, oversight, and behavioral evaluation.