Who’s Organizing the Agents?

The researchers at Duke University gave their AI agents a wildfire to deal with. Working in a simulation, some agents would scout with drones while others transported firefighters or fought the fire. With as many as 50 agents on a team, there was plenty of capacity to divide up the work. There were also plenty of opportunities for one agent to be waiting on another. The researchers wanted to know how much more these teams could accomplish if they changed the way the work was organized.

Across 25 missions using eight language models, they tested a framework called ORCH. It grouped agents according to the work each mission required. Drones could scout several areas at once, while transporting crews depended on information gathered earlier. Manager agents used reports from the team to revise assignments and decide when work could move to the next stage.

The researchers tried both team structures designed by people and structures generated by AI. Compared with four other frameworks, the human-designed ORCH organizations achieved roughly 64% higher final mission scores in the study’s weighted averages. The AI-designed organizations improved scores by about 44%.

Those gains reflect the combined effects of structure, planning, and feedback in a simulation. They don’t establish what improvement a company should expect. The coordination problem is familiar, though. Some work can happen independently, while other work has to wait for information or a decision from elsewhere in the process.

Consider a company using agents to onboard suppliers. One agent could check the submitted documents while another verifies supplier information. Both could finish faster than the people doing those checks today. Getting the supplier approved, though, would still depend on what happens after the checks are done.

Suppose one raises a concern. The agent reports it correctly and completes its assignment, but someone still has to decide whether the company can proceed. Other agents may need to pause their work until that decision is made. The check is finished, and its time savings can be measured, while the supplier is still waiting.

An employee handling this process might already know which concerns they can resolve themselves and which need another person’s approval. They may also know who to contact when an answer hasn’t come back, or which later steps can safely continue in the meantime. Some of that knowledge may be documented. Some may have accumulated through years of dealing with exceptions.

A supervising agent could take on some of those responsibilities, too, provided the company has worked out what decisions it can make and when it needs a person’s judgment. It would also need to make sure the other agents act on the decision. A report saying that a concern has been resolved won’t help much if the next agent is still waiting for permission to continue.

For an executive considering the investment, this affects both the savings and the effort required to achieve them. Faster checks have value, but they may do little to shorten onboarding if approvals still wait in the same queue. Improving how decisions move through the process could produce a larger gain. The business case needs to account for the work people will still do after the agents report that they’re finished.

The people coordinating the process today can help identify that work. A case that went through without a problem may tell you very little about their contribution. Ask them to walk through one that needed repeated clarification or an exception to the usual procedure, and the decisions an agent team would need to handle become easier to see.

If the business case includes reducing staff, those conversations need to happen before a position is counted as a saving. Its listed tasks might all be automated while much of its coordination remains unaccounted for. That work could return to the people who remain, one unresolved case at a time. The agents could keep completing their assignments successfully while those employees spend more of their day dealing with what happens afterward.

Algorithm and Blues publishes Sundays.

Reference

1. Ji, Z., Hyun, J., & Chen, B. (2026). ORCH: Organizational Principles Enable Collective Intelligence in Embodied AI. arXiv preprint, September 10.