Beyond Text-Passing: Shared Cognitive Substrates for Multi-Agent LLM Coordination

Current multi-agent LLM systems rely on natural language as the sole medium for inter-agent communication, treating coordination as a message-passing problem. We argue this text-passing bottleneck is a structural limitation—not merely an engineering inconvenience—that fundamentally constrains consistency, efficiency, and auditability. We propose a paradigm shift: from message-passing to shared cognitive substrates, where agents coordinate through three explicit shared primitives: (1) a Shared World Model maintaining typed state with invariants, (2) a Shared Causal Graph representing explicit dependencies and attribution paths, and (3) a Shared Energy Pool enabling resource-aware arbitration under budget constraints. We introduce Baton Arbitration as a coordination mechanism and propose Structural Continuity as an evaluation criterion for system stability. We present this as a blueprint intended to guide the design of substrate-native multi-agent systems.