Policy Inheritance and Override Patterns

As AI agent deployments grow, managing individual policies for every agent becomes impractical. Policy inheritance lets you define shared rules at a higher level and allow individual agents or teams to customize behavior within controlled boundaries. Override patterns determine what can be changed and what is fixed.

Base Policies

A base policy defines the rules that apply to all agents in the organization. These typically include hard security constraints: no data exfiltration, no execution of arbitrary code, no access to production databases without approval. Base policies represent non-negotiable safety requirements.

name: "org-base-policy"
default_effect: "deny"
rules:
  - action: "network.egress"
    destinations: ["*.internal.company.com"]
    effect: "allow"
  - action: "network.egress"
    effect: "deny"

Team-Level Policies

Team policies inherit from the base policy and add rules specific to the team's function. A data science team might allow database read access. A customer support team might allow email sending. Team policies can add permissions but should not be able to remove base policy restrictions.

Agent-Level Overrides

Individual agents may need specific exceptions. An agent that monitors external APIs might need egress access to specific external domains. Agent-level overrides apply on top of team policies but are still constrained by the base policy.

Override Constraints

The critical design decision is what overrides are permitted. Two common approaches:

Additive only: Lower-level policies can add allow rules but cannot override deny rules from higher levels. This guarantees that base policy restrictions are always enforced.

Scoped override: Lower-level policies can override higher-level rules, but only within a declared scope. For example, a team policy might be allowed to override egress rules for specific domains but not override the code execution restriction.

Conflict Resolution

When multiple policy levels contain conflicting rules, the engine needs a deterministic resolution strategy. Authensor evaluates policies from most specific to least specific, with explicit deny taking precedence over allow at the same level.

Design your inheritance hierarchy before writing individual rules. The hierarchy is the architecture; the rules are the implementation.