IMACS Framework Decouples Organization, Coordination, and Collaboration in Multi-Agent LLM Systems

A recent publication on arXiv presents IMACS (Intelligent Multi-Agent Collaboration System), a framework designed to disentangle three interconnected aspects of multi-agent LLM systems: team composition (who is involved), alignment (how members synchronize), and the collaboration protocol (which algorithm integrates their efforts). By isolating these elements into distinct, interchangeable layers, IMACS facilitates the application and validation of traditional organizational theories—such as Belbin roles, Mintzberg coordination, and RACI accountability—as configurations. The framework integrates six established collaboration algorithms through a unified interface, allowing roles, coordination, and accountability to be independently adjusted. This separation permits controlled experiments with varying organizational roles while maintaining a consistent collaboration protocol. Furthermore, protocol selection becomes a learnable aspect via Adaptive Org Routing, a contextual-bandit meta-protocol. The paper can be found on arXiv with ID 2607.25446.