MCP roadmap bets on agent identity, unified HTTP transport, long-running tasks

The Model Context Protocol’s maintainers have published a refreshed roadmap organizing near-term work into five priority areas, several of which graduate from the previous roadmap’s ‘on the horizon’ list. The headline theme is that agentic workloads have outgrown simple request-response: MCP is building out primitives for long-running loops, server-pushed streamed results, and mid-flight steering. Concretely that means server-initiated events (webhooks and channels to end client polling), a cross-Working-Group review of how Tasks, transports, and triggers compose, and promoting the Tasks extension (SEP-2663) into the spec.

The most consequential shift is around identity. Today’s authorization model assumes a human clicking ‘approve’ in a browser, but a growing share of callers are agents running as cloud workloads with their own identity, acting for absent users or delegating scoped authority to sub-agents. The roadmap commits to standardized, standards-based agent identity — finalizing and driving adoption of DPoP, and defining an opinionated path through Workload Identity Federation, the ID-JAG grant, and token exchange, rather than pasted API keys and long-lived tokens. Maintainers also plan deeper engagement with IETF OAuth and WIMSE groups so the underlying standards keep pace. On the transport side, the 2026-07-28 release already made a remote MCP server behave like any other HTTP workload; the goal now is to unify on a single transport, including local servers speaking Streamable HTTP over stdio.

Two developer-facing efforts round it out. First, tightening tool-call results: today a response can return the same output in multiple forms with no way to know which a client will surface to the model, so the spec will standardize on one contract, plus a ‘progressive discovery’ mechanism so servers with hundreds of tools don’t force the model to pay for the entire surface up front. Second, better SDK ergonomics and conformance — increasingly important now that many developers build MCP clients and servers by pointing an agent at the libraries. Proposals (SEPs) that land inside these five areas get expedited maintainer review; those outside aren’t rejected but compete for scarce attention.