Dynamic Coalition Formation and Communication Pricing in Skill-Based Agentic AI Systems

An arXiv paper (2608.07532) presents a new approach for enhancing communication and selecting agents in multi-agent AI systems. The authors treat agent selection and communication as a cooperative game, introducing a task-conditioned net utility function that distinguishes between coalition-level expenses and individual agent activation costs. They suggest a greedy router and a marginal-value activation rule, extending the framework to optimize communication edges with associated costs. Additionally, they utilize estimated Shapley values to determine which agents to contact during execution. The study links the issue to submodular maximization, offering two limited guarantees: a curvature-refined bound for a specific monotone case and a tighter bound. This work tackles inefficiencies in existing architectures that either predefine communication or permit unrestricted broadcasting, which can result in higher token costs, latency, redundancy, and error propagation. The paper can be found on arXiv with the identifier 2608.07532.