We introduce SPAR (Scalable Prioritized Agent Routing), a novel communication protocol for coordinating tasks in a decentralized graph of autonomous AI agents (with optional human participants). SPAR evolves from the Social Online Routing (SOR) protocol originally developed for human social networks, generalizing its concepts to multi-agent systems. Like SOR, SPAR enables decentralized, peer-to-peer request propagation without centralized brokers, but it is redesigned for AI agents by incorporating dynamic priority queues at each node to manage tasks at scale. We adapt SOR’s I-Need, I-Have, and I-Thank message framework to agent networks and integrate SPAR with emerging agent communication standards, complementing protocols such as Anthropic’s Model Context Protocol (MCP) [3], Google’s Agent2Agent (A2A) [4], and the open Agent Network Protocol (ANP) [5]. An experimental evaluation in simulated agent networks demonstrates that SPAR’s prioritized, queue-based routing achieves significantly lower end-to-end delays and message overhead than baseline flooding approaches, without sacrificing success rates. We also discuss practical applications of SPAR, from collaborative problem-solving among AI assistants to human–AI teamwork, and consider limitations around security, privacy, and open-network scalability, along with future directions to improve and standardize SPAR in the broader AI agent community.