We develop a model that shows that an overconfident manager, who sometimes makes value-destroying investments, has a higher likelihood than a rational manager of being deliberately promoted to CEO under value-maximizing corporate governance. Moreover, a risk-averse CEO's overconfidence enhances firm value up to a point, but the effect is nonmonotonic and differs from that of lower risk aversion. Overconfident CEOs also underinvest in information production. The board fires both excessively diffident and excessively overconfident CEOs. Finally, Sarbanes-Oxley is predicted to improve the precision of information provided to investors, but to reduce project investment.