Meta-Belief Routing for Scalable Embodied LLM Teams

Communication is becoming a scalability bottleneck for embodied LLM agent teams. As teams grow, broadcasting exposes every agent to messages it did not need, consuming context tokens that could have driven reasoning. We argue that the missing piece is not smarter compression of communication but a \textit{meta-belief} over who currently holds task-relevant knowledge. We introduce \textbf{CREAC}, a training-free, plug-and-play router that builds and maintains this meta-belief solely by overhearing natural-language messages, and uses it to deliver each message only to agents whose state suggests they can act on it. When its own belief becomes too uncertain to route well, CREAC stops listening and starts asking, issuing targeted queries to the most likely knowers rather than falling back to broadcast. On C-WAH (3--6 agents) and CREW-Wildfire (15 heterogeneous agents), CREAC outperforms both planner-coupled systems and the strongest selective-routing baseline, while reducing token cost relative to planner-coupled systems by up to 65\%. A user study further shows that human collaborators rate CREAC-mediated teams as more trustworthy and efficient than broadcast and silent baselines. Across both benchmarks, under a cooperative premise in which agents do not deliberately withhold task-relevant knowledge, delivering each message to the agents who can act on it improves coordination at lower token cost.