MCPContextShare-data

The Model Context Protocol (MCP) enables seamless context sharing between AI agents, yet the performance implications of different sharing strategies remain unexplored. This paper presents the first comprehensive empirical evaluation of five context sharing strategies: Broadcast (BC), Publish-Subscribe (PS), Pull-on-Demand (PD), Hierarchical Caching (HC), and Hybrid Adaptive (HA) in multi-agent MCP systems. We systematically evaluate these strategies across three dimensions: performance (latency, throughput), resource overhead (memory, network), and consistency (staleness, conflicts) using realistic workloads from the MS MARCO dataset. Our experiments span 758 configurations varying agent count (5-100), context sizes (100-2000 tokens), workload types, and access patterns. Results show that no single strategy dominates across all scenarios: Hierarchical Caching excels as the optimal choice for production systems, achieving 106,388 ops\/s mean throughput (490\u00d7faster than Broadcast) with 0.004ms p95 latency and 264ms acceptable staleness; Pull-on-Demand achieves peak throughput but exhibits critical 11-second staleness; Hybrid Adaptive shows unpredictable performance and 49% degradation at scale; Broadcast provides perfect consistency at low throughput (217 ops\/s). We provide empirically-grounded decision guidelines mapping workload characteristics to optimal strategies. Our open-source benchmark and complete experimental data enable reproducible research and inform the design of production multi-agent systems.