On October 1, Progress Software announced new capabilities in Progress Agentic RAG, its retrieval platform for agents and LLMs. The headline items are a native Microsoft Teams app that returns source-cited answers inside Teams, a Smart Agent that performs autonomous multistep retrieval across its Knowledge Box data, MCP-connected applications and web sources, and a WordPress plugin on the ingestion side. The company frames it as extending across delivery, retrieval, and content ingestion without requiring teams to rebuild existing deployments.
Strip the vendor framing and the shape is worth paying attention to, because it's the shape most internal knowledge stacks are converging on.
Three planes, not one pipeline
The old RAG diagram — chunk, embed, top-k, stuff the prompt — is being replaced by a split into ingestion, retrieval, and delivery, each scaling independently.
Ingestion is where the ugly work lives, and shipping a CMS plugin is an admission that content sources want to push into the index rather than be crawled. That's the right direction: push-based ingestion with source-side change events gives you freshness and deletion propagation, which polling crawlers handle badly.
Retrieval is where "agentic" means something concrete. A planner issuing several dependent queries across a governed corpus plus live MCP tool calls is a different cost and latency profile than one vector lookup. You are now paying for multiple model turns per answer, and you need per-hop tracing or debugging becomes guesswork. The tradeoff is real: multistep retrieval reliably beats single-shot on decomposable questions and reliably burns tokens on questions that didn't need it. Route accordingly.
Delivery into Teams is the least technically interesting and the most consequential. Answers that don't live where work happens don't get used, and citation surfaces are what make an answer auditable rather than merely plausible.
The part that isn't finished
The MCP maintainers published an updated roadmap on August 22, authored by lead maintainers David Soria Parra and Den Delimarsky, with priority areas including HTTP-native transport unification and hardening, agentic messaging primitives, improved primitives, SDK developer experience, and — the one that matters here — agent identity and enterprise-ready security. The current specification version is 2026-07-28, shipped alongside SDKs.
That ordering tells you where the risk sits. When an agent fans out across Knowledge Box content, MCP-connected business systems, and the open web, something has to decide whether *this user's* agent may read *that* record, and carry that decision across every hop. Today most teams solve it by pre-filtering the index per principal, which doesn't extend to live tool calls, or by giving the connector a service account, which quietly flattens your permission model.
If you're building this: treat retrieval as the easy half. Design the authorization path before the ranking function, and log every hop with the identity it ran under.