Who Watches the Watchmen? Safe AI-Agent Failover Via MCP and CRDs — Seoul 2026 | Agentic AI Foundation