Orchestrating WASM-based MCP Tool Runtimes for AI Agents across Edge-Cloud Continuum

초록

While the Model Context Protocol (MCP) standardizes logical agent-tool interactions, current container-based execution models fail within the Edge-Cloud Continuum, where edge devices are particularly resource-constrained. Since persistent deployment is infeasible on the edge, the reliance on heavy containers for on-demand execution causes prohibitive cold-start latencies, disrupting real-time agentic workflows. To bridge this gap, we propose an orchestration framework that decouples tool semantics from execution infrastructure, informed by our thorough analysis of MCP ecosystem heterogeneity. Based on these insights, we introduce the Agent Tool Dispatcher, a middleware that exposes detailed execution metrics to facilitate precise workload placement across Device, Edge, and Cloud tiers. Complementing this, we present WasmMCP, a WebAssembly-based runtime achieving architecture independence and rapid cold starts, leveraging Wasm’s sub-millisecond module instantiation to deliver end-to-end startup in tens of milliseconds to enable efficient deployment of MCP tools across constrained heterogeneous devices. Extensive evaluation using realistic agent scenarios demonstrates that WasmMCP achieves 10-21× faster cold starts compared to container-based baselines. Furthermore, our dispatcher achieves zero constraint violations while maintaining competitive end-to-end latency, ensuring seamless and scalable agent operations across edge-cloud continuum.

키워드