MCPステートレス仕様でAIエージェントインフラをスケーリングScaling AI Agent Infrastructure with the MCP Stateless updates

The Model Context Protocol (MCP) has released a revised specification, dated July 28, 2026, that replaces its earlier stateful requirements with a fully stateless core. The change matters because it removes one of the main obstacles to running AI agents at scale, allowing teams to rely on established cloud-native patterns such as horizontal scaling, serverless deployment, and standard round-robin load balancing rather than building custom infrastructure to preserve session state.

MCP is an open standard for connecting AI models and agents to external tools, data, and services through a consistent interface. First introduced by Anthropic in late 2024 and subsequently adopted across much of the industry, it defines how a client—often an agent powered by a large language model—discovers and calls the capabilities exposed by an MCP server. Google's developer materials, which cover this update in the context of the Gemini ecosystem, position the stateless revision as a foundational step for production-grade agent deployments.

The technical heart of the change is the move away from persistent, connection-bound session state. Under the older design, an MCP server generally needed to maintain context tied to a particular client connection for the life of an interaction. That constraint made scaling awkward: because any given request had to return to the specific server instance holding its session, operators typically relied on sticky sessions, session affinity, or shared state stores to keep things consistent. Those techniques work, but they complicate deployment, reduce fault tolerance, and sit poorly with elastic, ephemeral compute.

By defining a stateless core, the July 2026 specification lets any server instance handle any request. That property is what makes ordinary round-robin load balancing viable, since a load balancer can distribute traffic across interchangeable instances without tracking which instance owns a session. It also aligns MCP with serverless platforms—such as Cloud Run, AWS Lambda, or similar function-based runtimes—where instances are spun up and torn down on demand and are not expected to hold long-lived state. In practice this should make horizontal scaling, where capacity is added by running more identical instances, considerably simpler to operate.

The source material indicates the specification also introduces standardized mechanisms alongside the stateless model, though the specifics of that standardization are not fully detailed in the available summary. Broadly, a stateless protocol tends to push any required context into each request or into an external, shared layer, so that individual servers remain interchangeable. Readers evaluating the update should consult the full specification for the exact behaviors, migration guidance, and any compatibility considerations for existing stateful deployments.

For context, MCP sits within a fast-moving landscape of agent-oriented standards and tooling. It is frequently discussed alongside Google's Agent2Agent (A2A) protocol, which addresses communication between agents rather than between an agent and its tools, and alongside framework-level efforts to orchestrate multi-step agent workflows. The two concerns are complementary: MCP standardizes tool access, while agent-to-agent protocols standardize collaboration. A stateless MCP core is likely to make it easier to combine these pieces in the same cloud-native environment, because tool servers can now scale on the same terms as the rest of a distributed system.

The practical implications are most visible for teams already invested in containerized or serverless architectures. Removing session affinity requirements reduces the operational surface area, can improve resilience when instances fail or restart, and may lower costs by allowing tighter, demand-driven scaling. It also lowers the barrier for smaller teams that lack the resources to build bespoke state-management layers around an agent backend.

As with any specification change, the benefits depend on adoption. Existing MCP servers built around stateful assumptions will need review, and some workloads that genuinely require long-lived context will still have to manage that state deliberately, typically outside the protocol itself. Servers that maintain streaming connections or incremental progress may need particular attention during migration, since those patterns lean on continuity that a stateless model does not assume.

Even so, the direction is clear. By making statelessness the default, the July 2026 MCP specification appears designed to let AI agent infrastructure scale using the same well-understood tools that already run modern web services, narrowing the gap between experimental agent prototypes and systems that must serve production traffic reliably.