The Model Context Protocol got a new roadmap on Friday, and the through-line is hard to miss: almost every priority on it exists because agents stopped being chatbots that answer in one turn and became processes that run for hours. Lead maintainers David Soria Parra and Den Delimarsky published five focus areas developed alongside the Core Maintainers, community maintainers and the protocol's Working Groups.
Three of the five — server-initiated events, result type improvements, agent identity — appeared only as distant possibilities on the previous roadmap. Their promotion is itself a marker of how fast deployment reality moved.
Statefulness Treated as a First-Class Problem
MCP has accumulated Tasks, subscriptions, listen and progress notifications piecemeal as long-horizon workloads outgrew a single call that returns once. The ambition now is less about adding primitives than making the existing ones compose, through a review spanning the Agents, Transports, and Triggers and Events Working Groups. Two concrete deliverables sit underneath that: webhooks and channels so a client can stop polling for a result that may be an hour away, and graduating the Tasks extension, tracked as SEP-2663, out of extension status into the specification itself.
Steering matters as much as waiting. The maintainers single out the need to redirect work mid-flight — a requirement with no natural home in a protocol designed around one request and one response.
Who Is Actually Calling?
The authorization section is the roadmap's sharpest admission. MCP's current design assumes a person clicking approve in a browser, which was a fair description of an interactive client and a poor one for a caller that is a cloud workload holding its own identity, acting for a user who is not present, and possibly delegating narrower authority further down to sub-agents.
Rather than invent a scheme, the maintainers are betting on standards already in flight: finalising Demonstrating Proof of Possession and pushing its adoption, then defining an opinionated route through Workload Identity Federation, the ID-JAG grant underpinning Enterprise-Managed Authorization, and standard token exchange. They also intend to keep working inside the IETF's OAuth and WIMSE groups so those specifications grow the building blocks agent delegation needs. The alternative the document names — pasted API keys and long-lived tokens — is what a great deal of production tooling runs on today.
The Cost of a Hundred Tools
The most immediately practical item is about context economics. A server exposing a hundred tools charges the model for that entire surface before the user has asked anything, and selection accuracy tends to degrade as the list lengthens. Progressive discovery would let a server publish a small entry point and unfold the rest of its catalogue only as the conversation narrows.
A quieter fix targets ambiguity in results. A tools/call response can carry identical output in more than one shape, leaving server authors guessing which form a given client will put in front of the model; a single clear contract is meant to remove the guesswork.
Plumbing, and a Queueing Policy
Transport work continues from the 2026-07-28 release, which made a remote MCP server operationally indistinguishable from any other HTTP service and let teams host one wherever they already run their APIs. The next step extends that to local servers speaking Streamable HTTP over stdio, collapsing what remains of the transport split so client and server development converge on one path.
SDK ergonomics and specification conformance form the fifth area, and the maintainers tie its urgency to a specific habit: developers increasingly build MCP clients and servers by pointing an agent at the libraries, which turns accurate documentation into a functional requirement rather than a courtesy.
The roadmap doubles as a queueing policy. Enhancement proposals falling inside the five areas get expedited review and the best odds of acceptance; anything outside is not rejected on principle but competes for maintainer attention the document openly calls scarce. Contributors are pointed at the relevant Working Group first, with SEP-2133 available for teams that want to experiment in a sandbox repository before filing formally.
