Call the Agent mode API and your client inherits a work item whose length it does not control. The loop plans, queries, reads what came back, and revises the plan before answering [2]. Your process sits on an open HTTP connection while that happens, taking Server-Sent Events [5]. The announcement does not say what a client should do when that stream dies during iteration three, and it points you to the documentation for details and availability [17]. Check that before you put the integration behind a proxy with a sixty-second idle timeout.
The cost arithmetic is short. Chat mode answers a factual lookup with a single query [3]. Agent mode is defined by not stopping there, exploring data across multiple iterative queries [2], so the floor is two queries per question and the post names no ceiling [18]. Move a dashboard's worth of traffic across and you should price the floor, then watch the tail, because a plan that re-plans is charming right up until it lands on your latency budget.
File reasoning runs inside the same loop [10]. That is the honest wiring, since retrieving the right passage from a slide deck is itself a multi-step task [8], but it also means document questions inherit the query pattern and the latency of a deep table question. Content search is the stated mitigation: it prepares and indexes the files so retrieval across large volumes is faster and better [12]. Treat it as a build step, and work out its refresh story before you point an agent at a volume that changes daily.
Two pieces here are good engineering. Permissions are inherited from Unity Catalog, so a user sees only the files they already have access to [11], which keeps the authorisation decision in one place instead of in your prompt. And conversations and messages can be pulled back through the API for programmatic monitoring [6]. That is the first thing I would use: log the plan and the citations, so an answer can be checked after somebody has already pasted it into a board deck. The API also returns visualizations and data tables alongside text [7], which means your custom surface either renders what it is handed or degrades to prose deliberately.
Databricks says Agent mode can act as a real-time research partner for open-ended questions that previously needed an analyst and a full research plan [16]. For that to transfer to your workspace, the semantics the analyst held in their head have to exist somewhere the loop can read: agent instructions, table metadata, or the query history that Genie Code mines when it drafts a baseline agent from your description and example questions [14][13]. Where a table is new and its query history is thin, the generator has less to copy, and you are writing the instructions yourself.
In my context this is worth wiring into an internal tool first, where the caller is a known employee and a slow answer is tolerable. The customer-facing portal is the deployment that needs the identity and reconnection answers the post leaves to the docs.
