What are guardrails in Agent design?

• A. Safety constraints that prevent unsafe actions
• B. A style guide for writing tool descriptions
• C. A logging standard for tool usage
• D. Safety constraints that prevent unsafe actions, require confirmations for risky steps, or restrict tool access.

Answer: (D)

Guardrails in agent design are safety boundaries that shape how an AI acts to prevent harm and keep actions under control. They do more than just block dangerous moves; they also guide the process with checks and limits. Specifically, guardrails can stop unsafe actions from occurring, require explicit confirmation before proceeding with high-risk steps, and restrict which tools or data the agent can access. This combination makes the system safer and more reliable because it reduces the chance of accidental or unauthorized actions.

For example, an agent might be prevented from executing destructive commands, must prompt a user before performing a risky operation, and is prevented from using privileged tools unless approved. While individual aspects like just blocking unsafe actions or focusing on logging can be helpful, including all three elements—preventing unsafe actions, requiring confirmations, and restricting access—offers the most robust protection.