When implementing inter-agent communication for a distributed agentic system running across multiple NVIDIA GPU nodes, which message routing pattern provides the best balance of reliability and performance?