Contracting a crowd of heterogeneous agents

Technical Summary: Contracting a Crowd of Heterogeneous Agents

Problem Formulation
The paper addresses the optimal contract design problem for a principal managing a large population of heterogeneous agents. Unlike standard principal-agent models involving a single agent or a homogeneous crowd, this framework incorporates network spillovers where an agent's effort affects not only their own output but also the productivity of others. These interactions are modeled via an interaction function $G: [0, 1]^2 \to \mathbb{R}$, where $G(u, v)$ quantifies how the output of an agent of type $u$ is influenced by the effort of an agent of type $v$.

The setting is a continuous-time, linear-quadratic moral hazard model. The principal offers contracts consisting of a terminal payment and a continuous payment stream to incentivize agents who choose effort levels to maximize their own utility (balancing compensation against effort costs) while facing idiosyncratic noise. The principal aims to maximize the expected aggregate output net of total compensation, subject to the agents' participation constraints and incentive compatibility (Nash equilibrium) conditions.

Methodology
The authors employ a two-pronged approach, solving both the finite $N$-agent problem and its continuum limit, and then establishing the convergence between them.

1. Continuum Formulation via Labeling: To handle the technical difficulties of uncountably many interacting Brownian motions, the authors adapt the "labeling" technique from Lacker and Soret [LS22]. Instead of treating agents as distinct entities, they introduce a type variable $U \in [0, 1]$ as part of the state. The population distribution is defined as the joint law of the agent's type and output, $\mu \in \mathcal{P}_2([0, 1] \times C)$. This transforms the heterogeneous system into a standard mean-field game played by a representative agent against the population distribution.
2. Reduction to Mean-Field Control: Following the framework of Élie, Mastrolia, and Possamaï [ÉMP19], the principal's problem is reduced to a mean-field control problem. By parameterizing admissible contracts using Backward Stochastic Differential Equations (BSDEs), the authors characterize the agent's optimal effort and continuation value. This allows the principal's problem to be solved via a dynamic programming approach, leading to a Hamilton-Jacobi-Bellman (HJB) equation on the space of probability measures $\mathcal{P}_2(\mathbb{R}^3)$.
3. Verification and Explicit Solutions: A verification lemma (Lemma 2) is established to construct optimal contracts from classical solutions of the HJB equation. In the linear-quadratic case, the authors derive explicit solutions for both the finite and continuum problems. The continuum solution relies on solving an infinite-dimensional linear ordinary differential equation (ODE) system.
4. Convergence Analysis: The paper proves that as $N \to \infty$, the equilibrium efforts, optimal contracts, and the principal's value in the finite-agent model converge to their continuum counterparts. Crucially, it demonstrates that the continuum optimal contract, when evaluated on a finite grid of agent types, yields an admissible contract for the $N$-agent problem that is $O(1/N)$-optimal.

Key Results

• Explicit Optimal Contracts:

  • Finite $N$: The optimal contract for agent $i$ is linear in the output path, with a slope determined by a vector $\bar{Q}^N(t)$ solving a linear ODE system involving the interaction matrix $G_{i,j} = G(i/N, j/N)$.
  • Continuum: The optimal contract for an agent of type $u$ is $\xi^* = R_a(U) - \int_0^T (\frac{1}{2}Q(t, U)^2 + Q(t, U) \mathbb{E}[G(U, v)X_t]) dt + \int_0^T Q(t, U) dX_t$. Here, $Q(t, u)$ is the solution to the infinite-dimensional ODE $\dot{Q}(t, u) = -\int_0^1 G(v, u)Q(t, v) dv$ with terminal condition $Q(T, u) = 1$.
  • Interpretation: The function $Q(t, u)$ serves a dual role: it is the slope of the performance-based component of the contract and the equilibrium effort level induced in an agent of type $u$.

• Comparative Statics and Influence:

  • Incentive Targeting: The paper proves that agents with greater "outward influence" (defined by the integral of $G(v, u)$ over $v$) are incentivized to exert higher effort. Specifically, if agent type $u_1$ is less influential than $u_2$, then $\alpha^*(t, u_1) \leq \alpha^*(t, u_2)$.
  • Value Density: The principal's value can be decomposed by source type. The "source value density" shows that more influential agents contribute more to the principal's value, justifying steeper incentives for them.
  • Spectral Analysis: In symmetric interaction settings, the optimal effort is expressed as a series expansion over the eigenfunctions of the interaction operator. Modes with large positive eigenvalues (strong amplification) and large coefficients (contribution to aggregate output) receive the strongest incentives.

• Stability: The optimal contracts and equilibrium efforts are shown to be stable under perturbations of the interaction function $G$. Small changes in $G$ result in bounded changes in the optimal effort and the distribution of contracts.

• Convergence Rates:

  • The difference between the finite-agent equilibrium effort and the continuum effort is $O(1/N)$.
  • The difference in the principal's value between the finite and continuum problems is $O(1/N)$.
  • The "loss" incurred by using the continuum contract in a finite $N$-agent setting is $O(1/N)$.

Significance and Claims
The paper claims to provide a scalable approximation for contracting problems involving large, interacting populations. By solving the continuum limit, the authors obtain a tractable solution that can be applied to large finite samples without solving the high-dimensional finite-agent problem directly.

The primary economic insight is that optimal incentives must account for network position. In environments with spillovers, the value of incentivizing an agent depends on their ability to propagate effort through the network. Consequently, the principal should assign steeper incentives to agents who are more influential, rather than treating all agents uniformly or based solely on individual productivity.

The methodology bridges the gap between mean-field game theory and principal-agent literature, extending results from homogeneous settings (e.g., Élie et al. [ÉMP19]) to heterogeneous networks. The paper emphasizes that while the continuum model is an approximation, it is quantitatively controlled, offering a practical tool for designing contracts in complex networked environments such as team-based firms, platform economies, and supply chains.