Eliciting Language Model Behaviors with Investigator Agents