Self-Improving Agent • Multi-Model Deliberation