Actions Speak Louder than Words: Agent Decisions Reveal Implicit Biases in Language Models