Policy Enforcement

Consonant intercepts every agent call before it reaches the agent. It evaluates the request against your defined OPA policies. If the policy denies the action, the agent is never invoked.

Writing Policies

Policies are written in Rego. Consonant provides the following input to your policies:

{
  "agent": "stripe-agent",
  "capability": "refund_charge",
  "parameters": {
    "amount": 500,
    "user_id": "cust_123"
  },
  "user": {
    "role": "support_agent"
  }
}

Example: Limit Refund Amount

package consonant.policies

default allow = false

# Allow refunds under $100 automatically
allow {
    input.capability == "refund_charge"
    input.parameters.amount < 100
}

# Require approval for refunds over $100
require_approval {
    input.capability == "refund_charge"
    input.parameters.amount >= 100
}

Human-in-the-Loop

When a policy returns require_approval, the workflow pauses. An operator must approve the action via the Consonant UI or API before it proceeds.

Global vs. Agent Policies

You can apply policies at different levels:

• Global: Apply to all agents (e.g., "No PII in logs")
• Namespace: Apply to specific teams (e.g., "Finance team only")
• Agent: Specific constraints (e.g., "Read-only DB access")