Your AI Agent Isn't Done Until the Database Says So

A practical way to think about reliable tool-using agents: narrow actions, durable state, external verification, and failure paths that stop safely.