Background: Generative AI agents will need to work together, which requires monitoring and managing their performance. Objectives: The chief objective of this paper is to understand the joint design choice of the number of agents and their rewards. Methods: We study this problem in a theoretical framework of optimal incentives, where a system designer (principal) selects the environment in which multiple autonomous decentralized AI agents work together. These agents respond to incentives, such as rewards and penalties. We first consider a principal who selects the size of the agent team in addition to their incentives. Results: We prove a general result that the optimal team size will vary with the parameters of the environment, but the optimal incentives will not. This invariance property shows that agents should have different-sized teams on work projects rather than differing financial incentives. Conclusions: We show these results are robust to a more general framework, where the principal employs a supervisory AI agent to manage the tasks of the underlying AI team. Finally, we propose different levels of quality for the supervisory and worker agents, and find that it is efficient to match the best supervisors with the best worker agents.

This paper proposes a practical non-monetary mechanism that induces the efficient solution to the optimal rate control problem, where each client optimizes its request arrival rate to maximize its own net utility individually, and at the Nash Equilibrium the total net utility of the system is also maximized. Existing mechanisms typically rely on monetary exchange which requires additional infrastructure that is not always available. Instead, the proposed protocol is based on efficient delay allocation, where the server controls the delay experienced by each client through an intelligent scheduling policy. Specifically, we present an efficient delay allocation rule for the server to determine the target delay of each client. Then we propose a simple scheduling policy to achieve such delay allocation. Furthermore, we design a distributed rate control protocol for the system to converge to the Nash Equilibrium. The optimality of our mechanism is validated via extensive simulations on two representative systems against a baseline mechanism with FIFO scheduling and centralized rate control.