Guardrails
Rules or classifiers that constrain what an agent may do, say or access.
Guardrails are checks applied outside the model: input filters (prompt-injection detection), output validators (schema, policy classifiers), tool permission lists, spend and step limits, and sandboxes. They are the mechanism by which a team converts an unreliable model into a bounded system. Gartner cites 'inadequate risk controls' as one of three reasons it expects >40% of agentic projects to be cancelled by 2027.