DBWhisperer is a system that sits between an LLM data-analyst agent and the database it queries. Its starting point is an observation: agentic analytical workloads are unusually repetitive. An agent working through a question issues long sequences of related queries that recompute much of the same work, and, unlike a human-driven workload, it can expose its reasoning trace, giving the database a preview of queries that have not yet been issued. DBWhisperer exploits both signals. A reactive loop learns from the execution plans of queries already run, while a proactive loop reads the agent's streamed chain-of-thought to anticipate what is coming next; a planner then weighs the recommendations from both loops against a storage and compute budget before committing to actions such as materializing intermediate results, building session-local indexes, tuning database configuration, or rewriting incoming queries. The analyst agent is never modified and its reasoning is never constrained; its queries simply pass through DBWhisperer on the way to the database.