📄 中文摘要
在人工智能安全讨论中，“护栏”一词常被提及，但其具体含义却模糊不清。不同的工程师对护栏的理解各异，包括输出过滤、系统提示指令和主题限制等。这些措施虽然真实存在，但并不构成有效的政策执行或治理策略。有效的政策执行应在生产环境中具备可靠性、可审计性和有效性，而不仅仅是在演示中看似可行。文章强调了在实际应用中，如何确保人工智能代理遵循设定的规则和政策，以实现安全和合规的目标。
人工智能代理的政策执行：如何设定代理实际遵循的规则
📄 English Summary
Policy Enforcement for AI Agents: How to Set Rules Your Agents Actually Follow
The term 'guardrails' is frequently mentioned in AI safety discussions, yet its meaning remains vague. Different engineers interpret guardrails in various ways, including output filtering, system prompt instructions, and topic restrictions. While these measures are real, they do not constitute effective policy enforcement or governance strategies. Effective policy enforcement should be reliable, auditable, and effective in production environments, rather than merely plausible in demos. The article emphasizes how to ensure AI agents adhere to established rules and policies in practical applications to achieve safety and compliance.