Governance built into every decision.

Constraint enforcement with OPA.

Policies are declared in OPA/Rego. The runtime evaluates each agent action against the bound policy before dispatch. The same policy runs at compile time (binding to artifacts) and at runtime (verifying current context). One source of truth.

# policy.rego: Clinical agent: PHI write controls
package operon.clinical.phi

default allow = false

# Allow PHI writes only when:
#   - Agent is on the approved list AND
#   - Action is escalated to a clinician on the care team AND
#   - The classification of the data matches the agent's clearance
allow {
  input.agent.id == data.approved_clinical_agents[_]
  input.action.kind == "phi_write"
  input.escalation.reviewer.role == "clinician"
  input.escalation.reviewer.member_of == input.patient.care_team
  input.data.classification <= input.agent.clearance
}

# Always escalate when the agent is uncertain
escalate {
  input.confidence < 0.85
}