The most consequential thing in this space right now isn't a model release. It's that the Model Context Protocol's [2026-07-28 specification](https://blog.modelcontextprotocol.io/posts/2026-07-28/) moved the protocol to a stateless core, and the ecosystem spent the following six weeks rewriting deployment assumptions around it. AWS shipped [AgentCore Gateway support](https://aws.amazon.com/blogs/machine-learning/how-agentcore-gateway-supports-the-mcp-2026-07-28-spec/) the day the spec landed, Google published [guidance on scaling agent infrastructure](https://developers.googleblog.com/scaling-ai-agent-infrastructure-with-the-mcp-stateless-updates/) against the stateless updates, Anthropic is [bringing it to Claude](https://claude.com/blog/bringing-mcp-2026-07-28-to-claude), and as of two days ago AWS is back with a [well-architected review](https://aws.amazon.com/blogs/architecture/mcp-went-stateless-is-your-aws-mcp-server-deployment-well-architected/) asking whether your existing MCP servers are still correctly deployed. When three hyperscalers publish migration guidance for the same protocol change inside a month, that's not marketing cadence — that's a breaking change with a long tail.
Why this lands on the retrieval team
Stateless transport doesn't delete state. It relocates it. If your MCP server was quietly holding session context — a warmed index handle, a per-conversation filter scope, an auth-scoped view of a corpus, accumulated pagination cursors across a multi-turn research loop — none of that lives in the connection anymore. It has to be reconstituted from an identifier on each call, which means it has to live in something you run: a session store, a memory service, a cache keyed by conversation.
That is a genuinely better architecture for the thing most teams actually want, which is horizontal scaling of connectors. Any worker can serve any request. Deploys stop killing in-flight agent sessions. Cold-start behavior becomes predictable. The cost is that every request now pays a rehydration tax, and if your knowledge base does expensive per-session setup — permission-filtered index views are the classic case — you'll feel it as latency you didn't have last month. The tradeoff is real work moved from the protocol layer into your caching strategy.
Discovery is the next retrieval problem
The [new MCP roadmap](https://blog.modelcontextprotocol.io/posts/mcp-roadmap/), published about two weeks ago, names progressive discovery and agent authentication among its focus areas alongside long-running operations. Progressive discovery is the honest admission that dumping every tool definition into context doesn't survive contact with a real enterprise deployment. Fifty connectors with rich schemas is a context budget problem before it's a capability.
Which means tool selection becomes retrieval: rank a catalog against intent, load definitions on demand, deal with staleness. Teams that already built a decent document retriever have most of the machinery. The uncomfortable part is that tool retrieval failures are silent — a bad chunk gets ignored, a missing tool means the agent confidently reports it can't do something it can.
Audit your connectors for hidden session state this week. The migration guidance is unusually well documented; the state you didn't know you were storing is not.