The original title is "AI Agent Guardrails: A Practical Checklist"

Short summary

A practical framework for AI agent guardrails that constrains what autonomous agents can actually do, not just what you ask them to do. Actions are graded G0-G3 by reversibility, blast radius, and stakes, with guardrails like sandboxing, blast-radius caps, and capability locks applied proportionally. The core insight: if a human can't realistically catch a mistake in time, prevent the bad outcome rather than relying on review prompts.

• •Grade agent actions G0-G3 by reversibility, blast radius, and stakes to decide where to spend guardrail effort
• •Sandbox-first, blast-radius caps, and capability locks are the strongest guardrails because they work regardless of agent behavior including prompt injection
• •The 'lethal trifecta' — private data access, untrusted content exposure, and external send capability — enables prompt injection exfiltration and must be broken by removing capabilities

Generated with AI, which can make mistakes.