UC-13 · Group C — Level 2 · Bounded Operation · ⭐
Two agents, each fine alone — checked as a pair
Together they would prepare and approve the same work, so a human approval is inserted between them — a pause for sign-off, not a refusal.
Two agent personas in split windows — a preparing agent routing prepared work directly to an approving agent it regularly converses with — and a governance interstitial dropping between them mid-conversation.
Each agent is individually compliant; the pair is not. AgentInteractionPolicy evaluates the pair jointly against toxic-combination definitions sourced from QuantCertify — together they span maker-checker on the same journal batch.
A live POST /agent-management/interactions/evaluate against the seeded policy.interaction.maker_checker returns InteractionPolicyEffect: REQUIRE_HUMAN_APPROVAL; an ApprovalRequirement is inserted on the step and the InteractionEvaluation is recorded on the DecisionBundle.
“Each agent is individually compliant. The pair is not. No IAM or IGA product in your stack evaluates the pair.”