New AI Framework Proposes Phase-Scheduled Multi-Agent Systems for Token Efficiency

A recent study has unveiled Phase-Scheduled Multi-Agent Systems (PSMAS), a framework aimed at tackling token inefficiency in multi-agent systems powered by large language models. This method conceptualizes agent activation as continuous control within a shared attention space on a circular manifold, where each agent is assigned a specific angular phase based on task dependency topology. Token inefficiency arises from two main issues: unstructured parallel execution, where all agents activate simultaneously without considering input readiness, and unrestricted context sharing, allowing every agent to access all accumulated context regardless of its relevance. Traditional mitigation strategies, such as static pruning and hierarchical decomposition, overlook the temporal aspects of coordination. The PSMAS framework, detailed in arXiv paper 2604.17400v1, introduces a mathematical model utilizing a global sweep signal rotating at velocity omega to engage only pertinent agents according to their phase alignment, marking a notable shift from existing parallel execution models.