Documents your agents should be able to read
Connecting document stores to an agent is tempting to do all at once, because more context sounds like more capability. It is also the fastest way to give an agent access to things it should never quote and things that should never leave a shelf. Choosing what to connect first is worth a real decision.
The rule is the same as everywhere else in Lyrion: start narrow, widen on evidence.
Connect what answers the queue
The first documents to connect are the ones that answer the queue you started with. The refund policy, the product manual, the help centre, the standard procedures. Sources that are current, approved, and directly relevant to the work in front of the agent.
This keeps the agent's knowledge tight and correct. An agent that can only cite the current, approved sources is an agent that cannot confidently repeat last year's policy, which is exactly the failure mode you are trying to avoid.
Keep the stale and the sensitive out
The documents to leave disconnected are the out-of-date, the draft, and the confidential. An old policy the agent treats as current is a landmine. A draft it quotes as final is embarrassing. A confidential file it references in a customer reply is an incident.
Connecting a document store is granting the agent the right to speak from it. Treat it that way. If you would not want a sentence from a document appearing in a customer reply, that document does not belong in the agent's reach yet.
Freshness is a property, not a one-time check
A source that was current when you connected it goes stale on its own schedule. The connected set needs an owner and a review, so that when the policy changes, the agent's copy of it changes too. A synced source that silently drifts is worse than no source, because it is confidently wrong.
This is why Lyrion shows the sync state of every knowledge source. “Ready” and “syncing” are not decoration. They are the difference between an agent quoting today's policy and yesterday's.
Widen as the queue widens
As you hand over more queues, you connect the documents those queues need, one considered step at a time. The connected set grows with the work, not ahead of it, and every addition is a small decision with a clear reason.
Done this way, the agent's knowledge is always exactly as broad as its job and no broader. That is the same discipline as permissions and scope, applied to what the agent is allowed to read.
Where this tends to go wrong
The failure mode is almost never the agent inventing something wild. It is the small, plausible miss: a reply that is correct in general but wrong for this one account. That is exactly what the approval step and the log are for, and it is why we tell teams to read the log before they widen scope.