An emergency-ops agent must not call paging APIs unless a human confirms first. Which control best enforces that?

Guardrailing the emergency-ops agent so paging API calls require human confirmation is the control that enforces a policy that the agent must not page unless a human confirms first. Agent guardrails constrain high-impact tool use. Emergency notification and paging APIs can trigger citywide or large-group alerts; requiring a human-in-the-loop confirmation before those tool calls prevents autonomous false alarms, misuse, or prompt-driven abuse while still allowing the agent to prepare paging content for operator approval in municipal emergency operations.

Allowing the agent full autonomous paging rights to reduce on-call workload fails by concept because removing human confirmation violates the stated safety requirement and prioritizes convenience over controlled emergency messaging. Hard-coding every resident phone number into the system prompt for speed fails because embedding mass contact data in prompts creates confidentiality exposure and is not a confirmation control. Removing all tool-use logs so emergency actions cannot be second-guessed fails because it destroys accountability and forensics for high-impact civic actions rather than enforcing human approval.

Exam caveat: High-impact agent tools (paging, financial transfer, production change) often require human confirmation guardrails. Autonomy and log deletion are distractors when the stem mandates confirmation first.

Operational check: Configure the emergency-ops agent so paging API tools are gated on human approval, attempt an autonomous page (blocked), complete a confirmed page (allowed and logged), ensure resident phone books are not stored in the system prompt, and retain tool-use logs for after-action review.