AGTBRIEF
A new preprint argues the deciding resource is what each agent knows, not how many you deploy or how deep the hierarchy goes.
Picture a fleet of cooperating AI controllers tracking a moving target together: drones holding formation, or a cluster of LLM agents jointly monitoring a live system. When the job is shared that way, the engineering reflex is to throw more agents at it, or to nest them into a layered hierarchy. A new preprint from researchers at eCampus University, the University of Macerata, and V.N. Karazin Kharkiv National University says both reflexes usually miss the point.
Across robot swarms and collectives of LLM agents, the deciding resource is per-agent internal-model memory, not headcount and not depth. Run the authors' disturbance-rejection testbed, and a flat, homogeneous swarm with a matched internal model of the disturbance matches or beats a designed two-loop hierarchy at equal per-agent memory. Treat depth as a dynamical property, a recurrent internal model rather than a structural one, and the "irreducible causal floor" a prior preprint (cited as reference [9] by the authors) had claimed for flat systems dissolves.
The three resources, population width, per-agent memory, and the horizon an agent can predict across observation delay, do not trade off freely. The paper charts hard non-exchange boundaries on a width-by-memory map, and pins a residual error floor to observation delay times environment unpredictability, verified against the minimum-variance optimal controller.
Practitioners who pick architecture before sizing what each agent knows about the disturbance spend budget on the wrong lever. The full pipeline ships with the paper.