MCP Roadmap Delivers Stateless Spec as Agents Move Into Production

What changed

The Model Context Protocol has dropped the parts of its design that made it hard to run at scale. MCP is the open standard that lets AI agents connect to tools and data. Its maintainers published a 2026 roadmap on March 99, and the specification that came out of it, versioned 2026-07-28, was released on that date7. It is a change to the protocol's architecture, not a minor revision.

The roadmap set four priorities: transport scalability, agent communication, governance, and enterprise readiness1. It also changed how work gets planned. Instead of scheduling by release, working groups now drive the process. Specification Enhancement Proposals (SEPs) that fit one of the four priorities get faster review, and everything else faces a higher bar1. The roadmap's own explanation was that production deployments need different things than the early experiments that got MCP started1.

The biggest technical change is a stateless core. The new spec removes the initialize/initialized handshake, the Mcp-Session-Id header, and protocol-level sessions, which had forced operators to use sticky routing and shared session stores1. Every request now carries the protocol version, client details, and capabilities in its own metadata. A new server/discover method lets a client see what a server can do before opening a live connection1. The result is that MCP servers can sit behind ordinary round-robin load balancers or run on serverless platforms such as Cloud Run and Lambda without session pinning7.

A smaller core, with more pushed into extensions

The new spec also removes things from the core. Roots, Sampling, and Logging are deprecated in favor of tool parameters, direct calls to LLM provider APIs, and stderr or OpenTelemetry18. Tasks, the mechanism for long-running asynchronous work, now lives in a versioned extensions framework. A server returns a durable task ID, and the client checks on it or sends more input through it8. MCP Apps, which lets servers render interactive HTML inside a chat window in a sandboxed iframe, became the first official extension in January7.

To make these removals predictable, a new policy (SEP-2596) requires at least 12 months' notice before any feature is removed and keeps a public list of deprecated features8. The earliest that deprecated features can be removed is July 28, 20278. Under the old approach, breaking changes could arrive without warning8.

Authorization was tightened too. The release candidate makes OAuth 2.1 plus OIDC a requirement and adds an Enterprise-Managed Authorization (EMA) extension8. One implementer reports that EMA is now stable, that Anthropic, VS Code, Okta, Atlassian, Figma, and Linear have adopted it, and that Okta is the first supported identity provider7.

Why enterprises should care

The stateless redesign is mostly about the enterprise. Before it, scaling an MCP server across several instances usually meant workarounds. One write-up describes teams in 2025 using Redis to map sessions across Kubernetes pods just to keep requests reaching the right instance5. Removing protocol sessions gets rid of that problem instead of standardizing the workaround.

The roadmap names four enterprise gaps: audit trails that feed existing SIEM and APM tools, SSO-integrated authorization, gateway and proxy patterns, and configuration that moves cleanly between clients2. The more telling point is that the maintainers left this area deliberately loose. They asked enterprise practitioners to help define it, and they expect most of the work to arrive as extensions rather than additions to the core110.

Our reading is that this was the right call. A protocol that tried to build compliance logging, identity federation, and gateway behavior into its core would become hard to implement and slow to change. The maintainers are keeping the base thin and leaving room for a control plane that vendors and enterprises build around it. Commercial governance products are already filling that space. Mintive, formerly MintMCP, sells virtual MCP endpoints with SCIM-driven permissions and central audit logging14. SAP describes an MCP Gateway in its Integration Suite that manages the full lifecycle of MCP servers built from existing APIs15.

The cloud providers are lining up behind the spec. AWS's Bedrock AgentCore Gateway is reported to support the 2026-07-28 version as a managed gateway with unified logging and IAM7. One analysis notes that MCP now sits under the Linux Foundation and that AWS, Cloudflare, and Google have all made production commitments4.

Adoption numbers: same trend, different counts

Reports agree that adoption is growing quickly, but they measure it differently. Coverage from early 2026 cited about 97 million SDK downloads a month in February113. A September analysis said MCP's Tier 1 SDKs were close to half a billion monthly downloads by July 202620. Those figures may use different methods, so read them as a direction rather than a precise count. Even so, the client side has clearly converged. ChatGPT, Claude, Gemini, Copilot, VS Code, Cursor, and Microsoft's Agent Framework all reportedly shipped MCP clients by mid-2026, a level of vendor alignment on one tool protocol that is rare in the LLM era7.

Where A2A fits

The roadmap's agent-communication priority puts MCP closer to territory usually linked with Google's Agent2Agent (A2A) protocol. Most coverage describes the two as layers that work together. MCP connects an agent to tools and data, and A2A connects agents to other agents for discovery and delegation111216. Both are governed under the Linux Foundation's Agentic AI Foundation (AAIF)11.

A2A has matured on its own. It reached version 1.0 in April 2026 with signed Agent Cards for verifiable identity13. IBM's competing Agent Communication Protocol merged into it in August 2025, so the agent-to-agent layer consolidated instead of splitting13. The Linux Foundation reported more than 150 supporting organizations in April, including AWS, Google, IBM, Microsoft, Salesforce, SAP, and ServiceNow20.

SAP gives a concrete example of how the split works in practice. It uses MCP to give its Joule agents structured access to SAP business capabilities. For interoperability with outside vendors and third-party agents, it names A2A as the preferred approach and exposes Joule agents through an A2A-based Agent Gateway15. Other common descriptions follow the same pattern: an orchestrator hands work to specialist agents over A2A, and each specialist uses MCP for its own tools1114.

Reports differ on how committed some vendors are. A procurement-focused analysis from April said Microsoft and Anthropic had not committed to A2A and treated their alignment as the main uncertainty for adoption17. Other sources say Microsoft's Agent Framework release candidate, released in February, shipped native A2A support alongside MCP7, and list Microsoft among A2A supporters20. Our reading is that the concern about Microsoft is now out of date. Anthropic's position, described by one map only as an A2A "client"16, is still the less settled one.

The overlap may grow. One implementer says the AAIF's post-July priorities include A2A alignment: message formats that would let an MCP server act as a peer to other agents7. If that happens, the clean line of "MCP for tools, A2A for agents" could blur. Some guides already tell readers not to build MCP servers that coordinate agents11.

Watch out for unofficial timelines

Some coverage presents unofficial interpretations as if they were the official plan. One widely shared guide gives quarter-by-quarter dates, including A2A integration in Q3 and a verified MCP Registry with SLA commitments in Q43. The official roadmap is organized by priority, not calendar dates, and it speeds up proposals by theme rather than promising them by a deadline12. Early commentary also expected session state to move into external stores such as Redis9. The final spec went further and removed protocol sessions altogether1. Enterprises planning migrations should work from the spec, not from secondhand roadmaps.

What's next

The roadmap lists several items as "on the horizon": event-driven triggers, streamed and reference-based results for large payloads, and stronger security such as DPoP and Workload Identity Federation59. Server Cards, a metadata format served from a .well-known URL so registries and crawlers can find servers without connecting, are meant to address discovery210.

The more immediate task is migration. Every production server built for the 2025 stateful spec has until July 2027 before deprecated features can be removed78. One practitioner advises doing the OAuth work before the header changes, because a half-finished authorization layer breaks every client7.

MCP started as a convenience for developers. With this release it is trying to be infrastructure: stateless HTTP, standard identity, formal deprecation windows, and a governance process that fits enterprise expectations1. Its relationship with A2A is still being worked out. For organizations deploying autonomous agents, the practical result is that the base layer is now stable enough to build on, while the controls around it are still being developed.