We propose a framework to that learns to construct high-quality solutions efficiently in parallel for multi-agent combinatorial optimization problems