InfoQ Chinese published an edited transcript of a MAD Podcast conversation between Carnegie Mellon professor Andy Pavlo, now vice president of database research at ClickHouse, and host Matt Turck on how AI agents are reshaping database work. Pavlo argues that agents wiping production databases is an old problem resurfacing — permissioning and control mechanisms have existed for decades — and recommends isolating agents with restricted permissions and branch-based sandboxes rather than approving every action by hand. He also offers rough figures: off-the-shelf text-to-SQL tools reach about 60% accuracy, LLM-based database tuning gets roughly 80-85% of the way, and more than 60% of open-source database systems now show agent-authored commits in his dbdb.io tracking.