Guardrails

A guardrail is a check that sits at a boundary and decides whether something is allowed through. An agent has three such boundaries: what comes in (the input), what the model is allowed to call (its tools and actions), and what goes out (the output). A guardrail is a rule enforced at one of those points, designed into the places where an agent can do damage rather than bolted on at the end.

Treat an agent like a new hire on day one, with five controls: its own credentials, read-only access until write access is proven, scoped permissions, a full audit trail, and a kill switch. The five together limit what a mistaken or compromised agent can do. Untrusted input stays untrusted, every action is authorized on its own, and secrets are never stored in the client or the repo.