Dynamic Coalition Formation Optimizes Agentic AI Communication

Modern agentic AI systems, which combine multiple large language model agents with diverse skills, often suffer from inefficient communication. Current architectures typically either pre-define communication paths or allow full broadcasting, both of which lead to increased token costs, latency, redundancy, and error propagation as the number of agents and links grows. This research introduces a cooperative game model for agent selection and communication, optimizing based on task-conditioned net utility. It proposes a marginal-value activation rule and a greedy router, extending the model to include per-edge communication costs. Estimated Shapley values are used to predict which agents are most valuable to contact before and during execution. In synthetic experiments, this greedy routing approach achieved 99.5% of optimal utility while activating significantly fewer agents compared to full broadcast. The performance proved robust to varying activation costs and redundancy weights, though it declined under strong violations of submodularity or noisy value estimates. The framework offers a more efficient and scalable way to manage complex multi-agent AI interactions.