The most consequential thing happening in knowledge tooling right now isn't a model release. It's the slow, unglamorous migration of the connector layer to the `2026-07-28` Model Context Protocol revision — and the fact that implementations are only now catching up.
The spec itself shipped in late July, after a release candidate period, and the framing from nearly everyone writing about it is the same: statelessness and extensibility, with reworked authentication aimed at enterprise deployment. Google's developer blog has written about scaling agent infrastructure around the stateless updates; Microsoft shipped v2.0 of the official C# SDK; agentgateway added support in its 1.4 release about three weeks ago; the Agentic AI Foundation published a migration guide roughly two weeks ago, titled around the move "from local tool to distributed protocol." Security vendors — SecurityWeek, Stacklok, WorkOS — have all focused on the auth surface rather than the data surface.
That split is worth noticing, because the auth story has crowded out the retrieval story, and the retrieval story is the harder one.
What statelessness costs a retrieval server
If you built an MCP server that wraps an internal knowledge base, you very likely leaned on session state without calling it that. A session gave you somewhere to hang a cursor through a long result set, a cached embedding of the user's working query, a scratchpad of which documents the agent had already seen, and — critically — a resolved identity you only had to check once.
Take that away and every one of those becomes an explicit design decision. Pagination has to be encoded in opaque tokens the client hands back. Deduplication across turns moves to the client or into a side store you now operate. Permission filtering has to be re-evaluated per call against a token, which means your ACL check is now on the hot path of every retrieval, not amortized across a session. For a hybrid search backend doing BM25 plus vector recall plus a reranker, that's a real latency line item.
The upside is the one distributed systems people will recognize immediately: stateless servers scale horizontally, survive restarts, and sit behind ordinary load balancers and gateways. That is precisely why gateway projects were early to ship support — the stateless model is what makes a connector something an infrastructure team can actually run.
Who should care this week
If you maintain MCP servers over Confluence, SharePoint, a data warehouse, or a vector index, the migration guides now exist and the SDKs are moving. The work is not a version bump. Budget for re-expressing session-held retrieval state as explicit tokens, and for measuring what per-call authorization does to your p95.
If you're only consuming connectors, the near-term effect is a period of mixed-revision servers in the wild. Pin versions and expect inconsistent behavior at the edges until the long tail migrates.