Multi-agent AI systems are organizations
Argues that multi-agent AI systems are not merely technical architectures but organizations in the formal sense identified by a century of Carnegie-tradition organization science. Whenever multiple bounded agents pursue system-level goals, four universal problems arise: task division, task allocation, information provision, and reward (objective) distribution. Multi-agent AI systems face these problems by construction, not by analogy.
Reclassifies common AI failures through this lens: duplicated work is a task-allocation failure; incompatible outputs are joint task-division and information-provision failures; reward hacking and specification gaming are objective-distribution failures — the artificial-agent analogue of agency problems, arising from human under-specification rather than self-interest.
Three differences between human and artificial boundedness reshape the solution space. Motivation: artificial agents lack intrinsic preferences, shifting the problem from incentive design to objective specification — and making specification gaming the dominant failure mode. Information processing: AI agents gain less from repetition-based specialization and face lower inter-agent coordination costs, inverting the case for activity-based division of labour and making object-based or ensemble configurations more viable. Execution: rapid, high-fidelity execution converts design errors from recoverable friction into catastrophic cascades — artificial organizations require greater ex ante completeness of specification because they lack human improvisational correction.
On hybrid human-AI organizations: hybrids face the same four problems but must satisfy two sets of agent constraints simultaneously, which often pull in opposite directions. Identifies three cases — synergistic hybrids (complementary bounds), sub-dominant hybrids (contradictory requirements for speed and deliberation), and necessary hybrids (neither pure configuration is acceptable). The hybrid case is often the stable configuration, not a transitional state.
The paper's twin contributions: a practical call to design multi-agent AI as organization rather than architecture, directing attention from individual agent capability toward systemic coherence; and a theoretical opportunity — artificial organizations let researchers separate universal organizing problems from the historically contingent solutions human organizations developed, testing whether hierarchy, specialization, and peer review are solutions to organizing as such, or to organizing humans in particular.