An agent keeps calling a generic run_sql tool and occasionally issues an unbounded scan. Which change is most likely to fix this?

Adding more granular tools always improves an agent's accuracy, because the model has finer control over what it does.