Multi-Agent Contracts

A team of computer scientists including Paul Duetting, Tomer Ezra, and Michal Feldman has published a landmark paper titled 'Multi-Agent Contracts' in the Journal of the ACM, formally analyzing the complex problem of designing incentives for teams of AI agents. The research tackles a principal-agent model where a single entity (the principal) must motivate multiple agents to exert effort, with the collective outcome determined by a combinatorial reward function. The paper provides the first systematic computational analysis of optimal linear contracts for agent teams, focusing on reward functions within the complement-free hierarchy—a classification important in algorithmic game theory and economics.

The researchers' first major contribution is developing constant-factor approximation algorithms for finding near-optimal contracts when reward functions are submodular or XOS (a class including submodular functions), using value and demand oracles. However, their second, more striking result establishes fundamental limits: they prove that constant approximation is impossible for submodular functions even with full oracle access, and for the broader class of subadditive functions, they demonstrate an Ω(√n) approximation lower bound, where n is the number of agents. This degradation is notable because it applies even to functions very close to submodular, marking a significant departure from prior work in areas like combinatorial auctions. The findings imply inherent computational barriers to perfectly aligning and coordinating teams of AI agents using simple contractual frameworks, which has direct implications for multi-agent AI system design and automated mechanism creation.