A product manager deploys a procurement agent that can read contracts, draft purchase orders, and call a payment API, but leadership worries it might email vendors or spend above $5,000 without approval. The agent must stay helpful yet constrained. Which method should the manager use to implement agent guardrails that enforce safe boundaries?

The correct approach succeeds because effective guardrails combine preventive tool scoping with conditional autonomy, allowing the procurement agent to remain useful for reading contracts and preparing orders while blocking unauthorized channels like vendor email and pausing for human approval when spending exceeds $5,000. Comprehensive audit logging with dashboards is valuable for traceability, compliance investigation, and tuning thresholds, yet as a purely detective control it records violations after they occur rather than preventing unapproved emails or over-limit payments in the first place. Limiting the agent to read-only summaries and drafts without payment access improves safety through least privilege taken too far, but it violates the stated need to stay helpful by removing legitimate low-risk autonomy under $5,000 that leadership explicitly wanted to preserve with oversight. Setting a proper payment threshold while leaving vendor emailing unrestricted fixes the financial risk but fails on scope control, since autonomous external communication can create contractual commitments, confusion, or social-engineering exposure even without money moving. The exam caveat is to distinguish monitoring and manual minimization from enforced boundaries, choosing answers that actively constrain tools and escalate for approval rather than merely observe or entirely disable capability. As a practical check, test the deployment by attempting an over-limit purchase and an email send in a sandbox to confirm both are blocked pending sign-off while a compliant draft and small payment flow through cleanly.