The paper introduces ORCH, a method that organizes embodied AI agents into task-specific hierarchies using pooled interdependence for concurrent work and sequential interdependence for tasks with prerequisites. Tested on 25 wildfire-response missions with up to 50 heterogeneous agents and 8 language models, ORCH outperformed four prior multi-agent frameworks on outcome, efficiency, exploration, and resource use.
Human-designed ORCH improved final score by 63.97% and execution efficiency by 74.29% on average; LLM-generated versions improved by 43.63% and 52.53%. These gains held across missions and models. The work is interesting because it shows that organizational design, not just individual capability, drives collective intelligence in embodied AI.