Aligning LLM Agents by Learning Latent Preference from User Edits