In extreme cases, the coefficient can reverse direction entirely. A variable that appeared to *increase* $y$ in simple regression might *decrease* $y$ once confounders are controlled. This is Simpson's Paradox in regression form.