The Model Context Protocol (MCP) solves agent-to-tool integration, but it breaks Kubernetes scaling paradigms. When shifting MCP from local stdio to distributed clusters, platforms hit a wall: standard ingress controllers aggressively terminate long-lived Server-Sent Events (SSE), and stateless proxies fail to maintain the JSON-RPC state machines required by multi-agent loops.

This session engineers the "Agentic Data Plane." We deconstruct MCP’s architecture to expose why traditional L7 routing fails and provide a blueprint for a state-aware infrastructure layer that decouples AI hosts from tool servers.

Attendees will learn:

L7 Protocol Interception: Configuring Envoy to natively inspect MCP JSON-RPC payloads, enabling tool-specific rate-limiting.

Stateful SSE Fan-Out: Tuning TCP draining and keep-alives to support hours-long agent reasoning streams across ephemeral nodes without dropping context.

Transport Decoupling: Shifting from rigid stdio sidecars to scalable HTTP meshes.