Model Context Protocol Gets Its Biggest Overhaul Since Anthropic Launched It in November 2024

A Protocol Most People Have Never Heard Of Just Got a Lot More Important

The Model Context Protocol connects AI agents to the software companies already run: databases, internal tools, cloud services. Anthropic open-sourced it in November 2024. Twenty months later, it's become the default way AI agents talk to everything else, according to VentureBeat.

That quiet dominance now comes with a governance body attached. The update was released under the stewardship of the Agentic AI Foundation, a directed fund under the Linux Foundation, according to both VentureBeat and AAIF's own announcement. This is a meaningful shift. MCP started as an Anthropic project. Now it's run by neutral open-source infrastructure, the same model that governs Linux itself and Kubernetes.

What Actually Broke, and What Got Fixed

Under the old MCP design, an AI agent had to maintain a persistent connection to one specific server. Modern cloud computing doesn't work that way. Companies run fleets of interchangeable servers behind load balancers that spin up and down constantly. If the one server holding your agent's session died, the agent's work died with it.

Den Delimarsky, a lead maintainer of the protocol, described the old failure mode to VentureBeat bluntly: "Before, you needed to have a session store and manage session IDs, and if one of your compute pods went down, all of a sudden the requests would start failing."

The new release eliminates that requirement entirely by finalizing MCP's move to a fully stateless architecture, according to AAIF's own announcement. Companies can now run MCP servers behind ordinary load balancers using the same Kubernetes and cloud-native tools they already use for everything else. No special session-tracking infrastructure required.

David Soria Parra, MCP's co-creator and a lead maintainer at Anthropic, told VentureBeat the scale of the change justifies informal comparisons to a version 2.0: "Some people jokingly call it a v2, and I think in spirit that's accurate. It's probably the biggest change we've ever made to the protocol."

Guardrails Enterprises Actually Asked For

The update isn't just about scaling. It hardens the authentication model against what VentureBeat described as a known class of attacks, though the specific vulnerability class wasn't detailed in the announcement. It also introduces a formal 12-month deprecation policy, meaning any feature slated for removal now has a guaranteed year-long runway before it disappears, according to AAIF.

Companies building products on top of an open protocol need assurance the ground won't shift under them without warning. AAIF's announcement frames this directly: enterprises "cannot afford silent breakages or sudden architectural shifts." That's a reasonable ask from any company betting engineering resources on someone else's open standard, and it's the kind of stability corporate legal and procurement teams require before signing off on production deployments.

Two capabilities also graduated from experimental status into official protocol extensions: interactive server-rendered interfaces and long-running asynchronous tasks, according to VentureBeat. Mazin Gilbert, AAIF's executive director and a veteran of Google and AT&T, compared the statelessness shift to the architectural choices that made the modern web possible.

Snowflake Moves to Govern the Agents Themselves

The same week, Snowflake introduced Cortex AI Gateway, a separate but related product that governs and monitors AI agents as they interact with models, tools, MCP servers, and internal systems, according to SiliconANGLE. The gateway supports more than 100 MCP servers and centralizes access policies, authentication, and permissions.

The product builds on Snowflake's acquisition of Natoma Labs, an MCP governance startup Snowflake agreed to buy in May, according to SiliconANGLE. It also keeps a running log of exactly which agent did what, tracks token consumption and cost by team and workload, and lets finance and IT departments cap spending. This addresses the runaway-bill risk that comes with agents operating at scale.

Mayank Upadhyay, Snowflake's chief security and trust officer, said security has to be central to the shift toward agent interoperability: "Agent interoperability only works when enterprises can trust how agents from different platforms access data, invoke tools and take action on behalf of users."

Snowflake paired the launch with integrations from identity and security vendors including Aembit, Linx Security, SailPoint, Saviynt, and 1Password. Nancy Wang, 1Password's chief technology officer, said the integration gives third-party agents short-lived, task-scoped access instead of broad standing credentials. This addresses a common blind spot where it's unclear which agent accessed which data under whose authority.

What This Doesn't Solve

Neither the protocol update nor Snowflake's gateway addresses a harder question: as AI agents get more autonomy to act inside corporate systems, who's liable when one makes an expensive or destructive mistake. The governance tooling announced this week tracks and limits what agents do. It doesn't yet answer what happens after something goes wrong. That's likely the next fight as agentic AI moves from pilot programs into the kind of Fortune 500 production deployments AAIF says this update was built to enable.