AI Operations · Glossary
Policy
Component 2 of the Eight Components of an Interactive Agent. A Policy is an authored rule with a criticality tier that determines deterministic or probabilistic enforcement at runtime, expressed in natural language and operated against by the agent on every decision.
Policies are the atomic unit of the policy library. A Policy is authored by a domain expert, not engineered as code and not implied through prompt phrasing. Policies carry a criticality level (Low, Medium, High) that determines enforcement mode: High-criticality Policies execute deterministically, Low-criticality Policies operate probabilistically against the model's adherence. Policies have scope: Persistent Policies apply across all routines for the agent, Routine-Scoped Policies apply only within a specified routine, Continuous Policies evaluate on every turn. Policies are surfaced into the decision context through Top-K Policy Injection. Every Policy that fires on a decision is logged in the production trace. The policy library is the compounding asset.