ORCH Shows Org-Chart Design Beats Model Scale for 50-Agent Embodied Teams, Adding 63.97% to Mission Score

arXiv 2609.11737·medium signalXBlueskyLinkedInCopy link

Multi-agent systems are usually assembled with a fixed organizational structure regardless of what the task's coordination requirements actually are. ORCH builds task-specific hierarchies by combining pooled interdependence for concurrent work with sequential interdependence for prerequisite-governed work, and was tested on 25 wildfire-response missions with teams of up to 50 heterogeneous agents across eight LLMs. Human-designed ORCH organizations improved final score by 63.97% and execution efficiency by 74.29% over four prior frameworks, and LLM-generated organizations still gained 43.63% and 52.53%. Notably, collective performance was not monotonic in model scale, so the org structure carried more of the result than the backbone did.