loading
loading
Safety checks that prevent an agent from producing harmful or out-of-scope outputs.
Guardrails can be input filters (blocking dangerous prompts before they reach the model), output filters (scanning responses for policy violations), or structural constraints (limiting which tools an agent may call). They're a defense-in-depth layer — no single guardrail is foolproof, so production systems typically stack several. The trade-off is always between safety and false-positive rate.
Plainly
Think of Guardrails as the lock and rules that keep people and data safe. Safety checks that prevent an agent from producing harmful or out-of-scope outputs.
In practice
Use it when a workflow touches secrets, permissions, user data, payments, or untrusted input. In practice, define the owner, input, output, and failure mode before you rely on it.