If your business runs AI agents or has a team building them, today matters. The Model Context Protocol — the open standard that connects AI agents to the tools and data they need — is shipping its most significant spec update since launch.

The 2026-07-28 release introduces breaking changes that affect every AI developer working with MCP servers, and adds two new capabilities that could reshape how enterprise AI agents are built and deployed.

What Is MCP and Why Should Business Leaders Care?

MCP is the protocol that lets AI agents reach outside their model and actually do things: read documents, query databases, call APIs, run searches, update records. Without a standardized way for agents to connect to tools, every integration is a custom project. MCP changes that.

Since Anthropic open-sourced MCP in late 2024, it has become the default connectivity layer for AI agent work. Claude uses it. OpenAI’s agents use it. Most serious enterprise AI tooling now expects MCP-compatible servers. If you have a team building agents, they are almost certainly working with MCP.

What Changed Today

The Protocol Goes Stateless

The biggest change removes the initialize/initialized handshake that has defined how MCP sessions start since the beginning. Under the old model, every client-server connection had to establish a session — which meant AI agent traffic needed “sticky routing” to stay on the same server instance throughout a conversation.

That created real operational headaches: you couldn’t put a standard load balancer in front of MCP servers, horizontal scaling was awkward, and cloud deployments required extra infrastructure just to maintain session state.

The new spec drops all of that. Any MCP request can now land on any server instance. This is not a subtle change — it unlocks deployment patterns that were previously impractical and brings MCP infrastructure in line with how modern cloud services actually work.

Tasks Extension Ships Properly

Tasks have been in MCP’s experimental core since late 2025, but the July 28 spec promotes them to a formal extension with a redesigned API built around the stateless model.

Previously, when a client called a long-running tool, it blocked until the server responded. The new Tasks extension lets a server return a task handle instead of blocking, and the client polls with tasks/get — more like how real background jobs work. It also adds tasks/update for cases where the agent needs to give the server more information mid-task, and supports server-directed task lifecycles.

For enterprise deployments, this matters because it enables AI agents to kick off genuinely long-running operations — data processing jobs, document analysis pipelines, complex multi-step workflows — without holding a connection open the whole time.

MCP Apps: Agents That Ship Interfaces

The second official extension is MCP Apps, which lets MCP servers include sandboxed HTML interfaces that the AI client renders directly. This is a significant extension of scope: MCP servers could previously expose tools and data, but not UI. Now a server can provide an interactive form, a dashboard, a data table, or a configuration panel that appears inside the AI client alongside the conversation.

For businesses building internal AI tools, this opens up a category of agent-plus-interface experiences that previously required separate web apps.

Authorization Hardening

Six changes align MCP’s authorization model with OAuth 2.0 and OpenID Connect standards. This is the kind of update that doesn’t make headlines but matters enormously in enterprise environments: proper issuer validation, client type declarations, and tighter security boundaries around what agents can access.

If your security team has been nervous about AI agents accessing internal systems, the direction of the spec is clear. The protocol is maturing toward enterprise-grade access controls.

What This Means for Business

For executives overseeing AI deployment, the key takeaway is that MCP is growing up. The move to stateless architecture is the kind of foundational change that makes enterprise infrastructure teams breathe easier — it brings AI agent connectivity in line with how everything else in a cloud environment works.

For technical teams, the breaking changes are real. The handshake removal means existing MCP server implementations need updates. The Tasks redesign changes how long-running operations are handled. SDK maintainers have a ten-week window to validate against real workloads before the final spec is treated as stable.

For businesses evaluating whether to build on MCP or wait, today’s update is a signal to move. The stateless architecture, formal Tasks extension, and authorization hardening collectively address the three most common objections to production MCP deployment: scalability, async operation support, and security.

The businesses that build fluency with MCP now — understanding how agents connect to tools, how to instrument and govern those connections, and how to build workflows on top of them — will have a meaningful advantage as AI agents move from experiment to operational reality.

---

Enterprise DNA helps business leaders understand and deploy AI agents that work. If you’re thinking through how agentic AI fits into your operations, start with a conversation.