Org chart design beats bigger models for 50-agent teams

ORCH, a new hierarchy-design framework, lifted wildfire-response mission scores 63.97% over four prior approaches when humans built the structure, per a paper testing 50-agent coordination.

Collective performance wasn't monotonic in model scale across the eight models tested, and a stronger backbone often didn't rescue a weak structure. That's the direct stake for teams debugging multi-agent systems by upgrading models alone.

ORCH builds different hierarchies for different kinds of work inside one mission. Tasks that need several agents working at once get a pooled structure. Tasks that depend on an earlier step finishing first get a sequential chain of command.

The team tested ORCH on 25 wildfire-response missions with up to 50 heterogeneous agents drawn from eight different models. Human-built hierarchies also improved execution efficiency 74.29% over the four baseline frameworks. Hierarchies generated by an LLM instead of a human still beat the baselines, gaining 43.63% on score and 52.53% on efficiency.

The paper doesn't say whether the gap holds below 50 agents, or outside missions with clear prerequisite chains like wildfire response.