Model Context Protocol (MCP) is commonly presented as a lightweight protocol for stateless tool invocation. In production environments, however, MCP is increasingly deployed as a distributed system: scaled on Kubernetes, serving multiple concurrent agents, and operating under continuous capability change. In this setting, MCP behaves less like an integration layer and more like an agentic control plane.

This session explores MCP under production pressure. It highlights protocol stress points that emerge at scale - session instability under pod churn, replay and idempotency concerns, catalog inconsistency, backpressure, and multi-tenant policy enforcement - and explains why these issues arise naturally in high-availability, dynamic environments. Drawing on lessons from service meshes, event meshes, actor systems, and semantic data platforms, the talk outlines a layered architecture in which MCP provides minimal primitives while gateways handle routing, policy, consistency, and observability. Attendees will leave with a concise mental model and practical patterns for operating MCP safely at scale.