Enabling Self-Improving Agents to Learn at Test Time With Human-In-The-Loop Guidance