Design and place guardrails for an agent — input validation, output filtering, tool permission tiers, and human-approval gates. Use before an agent ships to users, when adding a risky tool, or when reviewing what an agent is allowed to do autonomously.
