Controlling Large Language Model Agents with Entropic Activation Steering