We propose a method to establish controllability using multi-agent deep reinforcement learning (MADRL) in sparse-reward environments. Advancements in MADRL have enabled agents to acquire sophisticated cooperative behavior by autonomously dividing tasks into subtasks. However, agents often learn behaviors that are reasonable but unacceptable from human requirements, ignore costly but socially required tasks, as well as struggle to adapt to environmental changes. Although we have already introduced a destination channel (DC) for this issue, we provide additional rewards when agents align with specified directives during training to enhance coordinated behavior. By providing DCs appropriately, agents can (1) learn actions that reflect human intentions, (2) avoid ignoring costly tasks/areas, and (3) prevent lazy agents. Managers can control coordination structures owing to environmental changes during the execution phases. We conducted experiments using an object-collection game to evaluate the proposed method against three baselines: the traditional DQN, implicit quantile network, and naive strategy-following distributed attentional actor architecture after conditional attention, which provide controllability only in simple environments. Our results demonstrate that the proposed method achieves efficient learning and controllability in complex environments by using indirect human instruction with DCs.