The headline features in the 2026-07-28 MCP spec — MRTR, header-based routing, an extensions framework — are downstream of one decision: the protocol core is now stateless. MCP went from a bidirectional stateful session to plain request/response, and every request is self-describing. That’s the change teams running MCP servers in production actually asked for.

Anyone who has tried to scale a stateful bidirectional protocol knows the tax. You need sticky sessions, connection affinity, and a load balancer that understands who’s talking to whom. Kill a pod mid-conversation and you’ve dropped state that was never yours to hold. Making any request landable on any instance behind round-robin is the difference between a demo and a fleet.

Two details matter more than they look:

• 🎯 Header-based routing — method and tool names ride in Mcp-Method and Mcp-Name, so a gateway authorizes and routes without parsing the JSON body. That’s real policy enforcement at the edge instead of deep in the app.
• 📊 Cacheable list results with deterministic order — tool catalogs cache, and upstream prompt caches stay stable across reconnects. In an agent loop where the tool manifest is re-sent every turn, prompt-cache stability is a direct latency and cost line item, not a nicety.

The tradeoff is honest: server-initiated sampling and elicitation now go through Multi Round-Trip Requests instead of an always-open stream, which is less elegant than true bidirectionality. The spec bet operability over interactivity, and for enterprise deployments that’s the right call.

The HN thread is already arguing about whether stateless-by-default was overdue or a capitulation. My read: the stateful model was the part everyone quietly worked around. Which of your agent tools actually needs a persistent session — or have you been paying for connection affinity you never used?