Guardrails are safety mechanisms that limit an AI agent within secure boundaries during execution. They include soft limits, like prompt recommendations for legal liabilities, and hard architectural limits. These structural guardrails ensure vertical agents cannot mistakenly use tools with business consequences. For instance, developers might separate database read from write access, requiring human approval for high-risk actions to ensure that AI agents never access restricted tools as a consequence of a hallucination.
Guardrails are also used when AI agents set expectations with employees regarding what they can and cannot help with. They can also provide legal feedback based on official company guidelines. Guardrails can involve rate limits to manage costs and limit the reach of unauthorized requests.
