The recent popularity of multi-agent conversation frameworks enabled by large language models such as AutoGen promises distributed problem-solving. However, these frameworks shed little light on how language emerges and how efficient the communication structure of different agents should be. In this work, we study communication among agents under various communication structures, where agents are transformer-based auto-encoders. Specifically, we study the emergence of language in stars, rings, wheels, randoms, and fully-connected structures. In each communication structure, a sender agent observes an image and sends a sequence of symbols to the receiver agent. The receiver agent reconstructs the image using the symbols. We evaluate the quality of the image using an MNIST classifier, which acts as the world model. Our results show that the star structure performs best, with agents achieving an F1 score of over 0.90 and a structural similarity index measure of 0.60. Agents in the star structure created a new language with minimum communication among themselves. The Levenshtein distance among the community members is less than 3 for 93 out of 100 test messages. In contrast, some agents in the ring structure did not learn the shared language. The Levenshtein distance in the ring structure is less than 3 for 40 out of 100 test messages.