This paper introduces CommNN, a neural model for multiagent communication that learns continuous communication protocols through backpropagation. The model consists of multiple agents that learn to communicate with each other during training, enabling them to collaborate on tasks. The agents use a continuous communication channel to exchange information, which is learned alongside their policies. The model is trained using reinforcement learning, with a policy gradient method that incorporates a baseline to stabilize training. The communication is continuous, allowing backpropagation and integration with standard single-agent RL algorithms. The model is simple, versatile, and scalable, allowing dynamic changes in the number and type of agents during runtime without exponential increases in state space or agent count. The model is applied to a variety of tasks, including a lever-pulling game, cooperative games in a maze environment, and a combat task. In the lever-pulling task, agents learn to communicate to pull different levers and maximize rewards. In the maze task, agents learn to navigate a traffic junction without collisions, with continuous communication significantly improving performance. In the combat task, agents learn to coordinate attacks against enemy bots, with continuous communication leading to higher win rates. The model also performs well on the bAbI tasks, a set of reasoning problems, outperforming some baselines while falling short of a more specialized model. The model's communication is interpreted in some cases, revealing simple strategies for task completion. The model's structure allows for different communication types, including continuous, discrete, and local communication. The continuous communication model is particularly effective in tasks with partial visibility, where agents must coordinate without full information. The model's ability to handle dynamic changes in the number and type of agents makes it suitable for a wide range of multiagent tasks. The paper also discusses related work, highlighting the novelty of the model in combining deep learning with reinforcement learning for multiagent communication.This paper introduces CommNN, a neural model for multiagent communication that learns continuous communication protocols through backpropagation. The model consists of multiple agents that learn to communicate with each other during training, enabling them to collaborate on tasks. The agents use a continuous communication channel to exchange information, which is learned alongside their policies. The model is trained using reinforcement learning, with a policy gradient method that incorporates a baseline to stabilize training. The communication is continuous, allowing backpropagation and integration with standard single-agent RL algorithms. The model is simple, versatile, and scalable, allowing dynamic changes in the number and type of agents during runtime without exponential increases in state space or agent count. The model is applied to a variety of tasks, including a lever-pulling game, cooperative games in a maze environment, and a combat task. In the lever-pulling task, agents learn to communicate to pull different levers and maximize rewards. In the maze task, agents learn to navigate a traffic junction without collisions, with continuous communication significantly improving performance. In the combat task, agents learn to coordinate attacks against enemy bots, with continuous communication leading to higher win rates. The model also performs well on the bAbI tasks, a set of reasoning problems, outperforming some baselines while falling short of a more specialized model. The model's communication is interpreted in some cases, revealing simple strategies for task completion. The model's structure allows for different communication types, including continuous, discrete, and local communication. The continuous communication model is particularly effective in tasks with partial visibility, where agents must coordinate without full information. The model's ability to handle dynamic changes in the number and type of agents makes it suitable for a wide range of multiagent tasks. The paper also discusses related work, highlighting the novelty of the model in combining deep learning with reinforcement learning for multiagent communication.