MCP maps its next spec: agent identity, Tasks, unified HTTP transport
The Model Context Protocol just published a fresh roadmap, and almost every line of it assumes the thing calling your tools is an autonomous agent, not a person sitting at a keyboard. Here's what the next specification is betting on.
MCP — the protocol that has become the default plumbing between AI models and external tools — laid out its next act on Friday, and it reads like a confession that the chat-era design no longer fits. Lead maintainers David Soria Parra and Den Delimarsky published the update alongside a refreshed roadmap page, organizing future specification work into five priority areas, each owned by core maintainers and a working group. Proposals inside those areas get expedited review; everything else waits in line, because maintainer attention — not ideas — is the scarce resource. If you need the baseline first, we explained the protocol in plain English in AI 101 — What is MCP?.
The first priority concedes that request-and-response was built for a world where a model answered and stopped. Agentic loops now run for hours, servers push streamed results back, and users want to steer mid-flight. The Tasks extension — which lets a server accept long-running work and report back later — gets matured toward formal spec status, joined by server-initiated events like webhooks so clients stop polling like it's 2024.
The quiet centerpiece is identity. Today, MCP authorization assumes a human approving access in a browser window. The maintainers want servers to recognize and trust agent identities instead: cloud workloads acting for users who aren't present, delegating narrow authority to sub-agents. Crucially, they're building on existing standards — proof-of-possession tokens, workload identity federation, enterprise-managed authorization — rather than pasted API keys and long-lived credentials, and coordinating with the IETF's OAuth and WIMSE groups. This is the piece that decides whether enterprises ever let agent fleets out of the sandbox.
The rest is consolidation and economics. Everything converges on Streamable HTTP — remote MCP servers are already just ordinary HTTP workloads since July's spec, and local ones are next. Ambiguous tool results get a single clear contract. And a new "progressive discovery" effort attacks a real cost problem: connect a model to a server with a hundred tools and it pays for the entire catalog before you've asked a single question. Servers would offer a small entry point that reveals depth only as the conversation narrows.
MCP won the adoption war; now it has to win the operations war. Whoever writes these rules is writing how millions of agent integrations authenticate, delegate, and scale. That's not changelog trivia — it's infrastructure policy.
What to watch: whether Tasks clears into the next formal spec release — and whether major MCP clients adopt agent identity the day it lands, because a standard nobody implements protects nobody.
Would protocol-level agent identity finally make you trust agents in production, or is it just more paperwork? Tell us in the comments.
Sources: Model Context Protocol Blog · MCP Roadmap Documentation · Hacker News discussion
