Google has published a technical walkthrough of the July 28, 2026 Model Context Protocol update, framing the change as a move from session-bound agent integrations toward more conventional cloud infrastructure.

The practical shift is in MCP's Streamable HTTP transport. Earlier HTTP deployments could depend on an initialization flow and an Mcp-Session-Id header, which meant later tool calls had to find the server instance holding that session state. Google's post argues that this made horizontal scaling harder for teams trying to run MCP servers behind ordinary load balancers or serverless platforms.

The new specification makes each request carry the information needed for routing and handling. The MCP docs show standard headers such as MCP-Protocol-Version, Mcp-Method, and Mcp-Name, with examples for tool calls and resource reads. Google says that design allows round-robin routing, cache controls, and trace-friendly HTTP behavior without tying a client to one container.

This matters because MCP has become a common interface layer between AI agents and external tools, but production systems need operational traits that local demos can ignore. A stateless request model should make it easier to scale tool servers, isolate failures, and fit MCP traffic into existing gateway and observability stacks.

There are still compatibility questions for clients and servers built around older session behavior. For developers building agent backends now, though, the update is a sign that MCP is being shaped less like a desktop integration protocol and more like shared AI infrastructure.