Current multi-agent AI systems have a fixed population. The number of agents, their roles, and their lifespans are set at design time. Garnier asks what happens when agents can be born, duplicated, specialized, and destroyed at runtime, with population dynamics governed by the system's own performance rather than external specification.
The framework models agent families as production sectors and computational resources as factors of production, with an orchestrator coordinating allocation. The dynamics are formalized using growth theory from economics. The central result: under specific parameter conditions, the system exhibits a Hopf bifurcation generating endogenous demographic cycles. Agent populations naturally oscillate between expansion and contraction without any external trigger. The system creates boom-and-bust cycles in its own workforce.
The mechanism is feedback between specialization and resource competition. When a task becomes important, agents specializing in that task are created. Their success attracts more resources. But the resource pool is finite, so other agent types are starved and die. The death of generalists creates gaps that specialists cannot fill, triggering demand for new generalists, which pulls resources away from specialists. The cycle repeats.
The through-claim is about the inevitability of oscillation in adaptive populations. A fixed agent population avoids cycles by construction — nothing changes. A dynamically adaptive population gains flexibility but inherits the instability that flexibility creates. The cycles are not bugs in the orchestration. They are structural consequences of coupling population dynamics to resource allocation under finite budgets. Any system that creates and destroys workers based on demand will oscillate, because creation in one sector is destruction in another, and the lag between creation and productivity ensures the response always overshoots.