Turns out AI agents work better when someone hands them an org chart.
Researchers built a framework called ORCH that assigns task-specific hierarchies to teams of AI agents, splitting work that can happen at the same time from work that has to happen in a fixed order. They tested it across 25 simulated wildfire-response missions covering reconnaissance, rescue, transport, resource management, containment and suppression, using teams of up to 50 agents built from eight different large language models. Compared with four existing multi-agent approaches, human-designed ORCH structures lifted mission scores by 64% on average and improved execution efficiency by 74%. Letting a language model design its own org chart instead of a human still helped, just less: a 44% score gain and a 53% efficiency gain.
The result cuts against the assumption that AI teamwork problems get solved by throwing bigger models at the coordination layer. The researchers found collective performance was not reliably tied to model scale - how the team was organized mattered more than which model ran it. That is a useful data point for anyone building multi-agent systems for physical tasks, from warehouse robots to drone swarms, where a single fixed coordination structure is still the default.
These are simulated wildfires, not real ones, so the gap between a tidy benchmark and smoke-filled reality is the part still worth doubting.