The most consequential thing in this space right now is not a model release. It is that the Model Context Protocol's 2026-07-28 specification removed the protocol-level session, and the ecosystem is still working through what that costs. The maintainers' updated roadmap, published August 22, puts it bluntly: a remote MCP server is now a normal HTTP workload, and the protocol increasingly leans on HTTP specifics — headers, status codes — to carry transport-level information.
If you build knowledge connectors, that sentence is the whole story.
What actually changed
The initialize/initialized handshake and the protocol-level session are gone from the core. Alongside that: header-based routing, cacheable list results, Multi Round-Trip Requests, a formal Extensions framework, Tasks, authorization hardening, a deprecation policy, and updated Tier 1 SDKs. The release candidate locked in May, giving SDK maintainers and client implementers roughly a ten-week validation window before the final spec published on July 28.
Three of these matter disproportionately for retrieval.
Cacheable list results. A serious enterprise connector exposes a large, mostly static catalog of tools and resources. Re-enumerating it on every connection was pure overhead against a corpus that changes hourly at most. Caching it is not a nicety; it is the difference between a connector that scales to thousands of concurrent agents and one that doesn't.
Statelessness. Any replica can serve any request. No sticky sessions, no session affinity in your load balancer, no warm in-process state to lose on a deploy. This is straightforwardly good for operators — and it relocates work rather than eliminating it. Everything a session implicitly held now needs an explicit home: token exchange results, pagination cursors over a large corpus, and above all per-user permission caches. Permission filtering is usually the expensive half of enterprise retrieval, and re-deriving an ACL set per request is not free. Budget for an external cache with a real invalidation story, or you will pay the identity provider on every query.
Tasks. Retrieval work that outlives a request timeout — reindexing, multi-hop deep search, large document ingestion — finally has a protocol-level shape instead of a bespoke polling endpoint per vendor.
The federation pull
This lands while enterprise search vendors push query-time federated and MCP connectors that read source systems at the moment of the query rather than serving pre-indexed snapshots. The appeal is obvious for fast-moving data — support tickets, tickets in flight, anything where a nightly crawl is wrong by morning — and permissions come out correct by construction because you are reading as the user.
The tradeoff is equally obvious and underdiscussed: you cannot rank across corpora you have not indexed. Federated results arrive as per-source lists with incomparable scores, and latency is now bounded by your slowest upstream. Most teams that ship this well end up hybrid — indexed for recall and cross-source ranking, federated for the handful of sources where staleness is unacceptable.
Actionable now: if you run a remote MCP server, audit for implicit session assumptions and read the deprecation policy. The migration window is the friendliest it will be.