This PhD will investigate how multimodal embodied AI agents can use language to communicate and coordinate in dynamic environments. A central goal is to develop benchmarks for evaluating communication, coordination, information sharing, and adaptation over long-running multi-agent interactions. The work will study how access to communication affects collaborative performance, including settings where agents have partial or differing information. It will also investigate mechanisms that allow agents to maintain, update, and use knowledge about other agents and the environment across extended interactions. Building on these benchmarks, the project will develop methods for improving communication as agents learn continually from interaction. The research will primarily use simulated embodied environments, enabling controlled evaluation across tasks, communication structures, and agent capabilities. The broader aim is to understand when language improves collaborative behaviour and how embodied agents can become more effective long-term partners for humans and other agents.