This paper introduces CommNN, a neural network model designed for multi-agent communication and cooperation. CommNN allows agents to learn continuous communication protocols during training, enhancing their performance in various tasks. The model consists of multiple agents, each controlled by a deep feed-forward network, and they communicate through a continuous vector channel. This channel enables agents to receive and transmit messages, which are learned during training. The model is trained using backpropagation, making it compatible with standard reinforcement learning algorithms. The paper demonstrates the effectiveness of CommNN through experiments on diverse tasks, including lever pulling, traffic junction control, combat scenarios, and the bAbI Q&A dataset. Results show that CommNN outperforms models without communication and those using discrete communication, indicating the importance of continuous communication in multi-agent cooperation. The model's ability to learn sparse and meaningful communication protocols is highlighted, and future work on handling heterogeneous agent types and scaling to large numbers of agents is discussed.This paper introduces CommNN, a neural network model designed for multi-agent communication and cooperation. CommNN allows agents to learn continuous communication protocols during training, enhancing their performance in various tasks. The model consists of multiple agents, each controlled by a deep feed-forward network, and they communicate through a continuous vector channel. This channel enables agents to receive and transmit messages, which are learned during training. The model is trained using backpropagation, making it compatible with standard reinforcement learning algorithms. The paper demonstrates the effectiveness of CommNN through experiments on diverse tasks, including lever pulling, traffic junction control, combat scenarios, and the bAbI Q&A dataset. Results show that CommNN outperforms models without communication and those using discrete communication, indicating the importance of continuous communication in multi-agent cooperation. The model's ability to learn sparse and meaningful communication protocols is highlighted, and future work on handling heterogeneous agent types and scaling to large numbers of agents is discussed.