The week's most telling development is not a model. On September 29 and 30, several outlets — AI News, TestingCatalog, Tao Media, Technobezz — reported that Meta shipped Muse for Small Business, extending its assistant with connectors into the systems small companies actually run on: accounting books, ad accounts, and general work tools, paired with goal-based skills. Meta's own help documentation on how Muse works with connectors went up about a week earlier.
Strip the consumer framing and this is the same architecture every enterprise knowledge team has been building for two years, aimed at a market that was never going to buy a knowledge platform. The interesting part is what it says about where the difficulty now sits.
The index is no longer the hard part
For a while, the defensible work in retrieval was ingestion and ranking: chunk boundaries, embedding choice, hybrid lexical-plus-vector scoring, reranking. Those are now close to commodity. Hyperframe Research's write-up of OpenSearchCon this week argues the same thing from the infrastructure side — search engines are settling into the role of the retrieval layer underneath agents rather than a destination product of their own.
What is not commodity: permission-correct, incrementally-synced access to twenty systems of record with different auth models, rate limits, and deletion semantics. A connector is not an ETL job. It carries an ACL model you have to mirror at query time, a change-feed you have to trust for freshness, and a tenancy boundary you cannot get wrong once. Most teams discover this when someone's retrieval surfaces a document from a channel they left in March.
What this means if you're building
Three tradeoffs worth being deliberate about:
Sync versus live fetch. Indexing gives you latency and ranking control; live calls at query time give you freshness and let the source system enforce permissions for you. Hybrid is usually right — index metadata and ACLs, fetch bodies on demand for the handful of documents that make it into context.
Structured versus text. Accounting ledgers and ad accounts are not documents. Embedding them destroys the thing that makes them useful. Route those to query generation against the actual API and keep the retriever out of it.
Who owns the auth boundary. Delegated per-user OAuth scales badly but fails safe. Service accounts scale well and fail loudly in the wrong direction. The rush of agent-security funding — TechCrunch reported Reco raising $55M on Monday into a crowded field — is a market response to teams having chosen the second option first.
The protocol layer is converging on this too: the July 2026 MCP specification and the roadmap published in August both point toward a stateless core with enterprise concerns handled as extensions. That is a bet that connectors become interchangeable plumbing. If it lands, the differentiator moves up a level — to what your system chooses to retrieve, and what it declines to.