interactive · 11 min

A Guardrail Layer

Objective: put a deterministic check between the policy’s chosen action and execution: allow only allow-listed tools, and reject unsafe arguments.

The policy — your Foundations MockLLM seam — may propose anything; this deterministic guard decides what actually runs. A blocked action returns a defined result the loop can handle; it never throws. The path check here illustrates the principle; production argument validation uses a schema or a dedicated library, not ad hoc substring scanning.

Best practice: allow-list, don’t deny-list; validate arguments structurally; a blocked action is a normal, logged outcome.

Next: Prompt Injection & Safe Tools