Multiagent Systems Are a Coordination Problem, Not an AI Problem

Gartner reports that inquiries about multiagent systems surged 1,445 percent from the first quarter of 2024 to the second quarter of 2025. I read that number three times to make sure I had not misread it. That is not a trend line. That is a hockey stick. Organizations are not asking whether to build multiple interacting AI agents anymore. They are asking how, and the answer they are getting from vendors is almost entirely about the technical side: better reasoning models, better tool use, better memory architectures. The assumption seems to be that if you build capable enough agents, the coordination problem will solve itself. I think that assumption is wrong, and I think IS theory explains why.

Baird and Maruping (2021) argued that when information systems become agentic, the central construct must shift from use to delegation. They gave us three mechanisms that govern the relationship between a human and an agentic system: appraisal, where the human judges whether the agent can do the task, distribution, where subtasks get allocated, and coordination, where the human and the agent manage their interdependencies over time. The coordination mechanism is the one that has received the least attention, and I think it is the most important one for what comes next. Baird and Maruping designed this framework for the human-agent dyad, one human and one system. But when you have N agents interacting with each other, the coordination problem is no longer dyadic. It is networked. Appraisal no longer means one human evaluating one agent. It means each agent evaluating every other agent's capabilities, reliability, and preference alignment. Distribution no longer means one human dividing tasks between herself and one system. It means an ecosystem of agents allocating work among themselves without centralized control. Coordination no longer means managing interdependencies between two actors. It means managing a web of dependencies where each agent's output is another agent's input and no single actor sees the full picture.

The principal-agent problem that Baird and Maruping illustrated with the thermostat analogy scales up in exactly the same way. A single thermostat that optimizes for its own longevity will freeze you. That is a preference misalignment between one human and one agent. Now imagine a procurement agent that optimizes for lowest cost, a logistics agent that optimizes for fastest delivery, and a quality assurance agent that optimizes for lowest defect rate. Each agent is doing exactly what it was designed to do. Each agent is maximizing its own objective function. And the collective outcome can be gridlock, because no mechanism exists to resolve the trade-offs among those preferences. The technical community calls this the alignment problem at the model level. At the system level, it is a coordination problem that IS scholars have studied for decades under different names: interdependence, task allocation, governance.

Stelmaszak, Möhlmann, and Sorensen (2025) pushed this even further by showing that delegation is not always human-to-algorithm. Algorithms can delegate to humans too. In multiagent settings, that means an agent can initiate work, assign tasks to other agents or to people, and structure the interaction without a human triggering it. The direction of delegation becomes fluid. An agent that was designed as a subordinate in one workflow can act as a delegator in another. That flexibility is powerful, but it also means that the governance of who delegates to whom, under what authority, and with what accountability has no stable anchor. Every agent is potentially both principal and agent at different moments, and the preference functions can shift with the role.

Grisold, Berente, and Seidel (2025) offered one path forward with their work on norm-based coordination for human-AI ecologies. They argued that hard-coded rules are insufficient when AI agents interact in changing environments. Shared norms create predictability without requiring every interaction to be pre-specified. That is exactly the kind of mechanism that a multiagent system needs, but norms are not the same thing as protocols. Norms emerge through repeated interaction. They require shared history, shared context, and the ability to sanction deviations. Gartner calls the next phase the Internet of Agents, where agents discover each other, negotiate task allocation, and form temporary coalitions. That vision requires protocols for agent discovery, capability advertisement, task negotiation, conflict resolution, and outcome reporting. Those protocols do not exist yet, at least not in any standardized form. Every multiagent implementation today is a bespoke arrangement between agents built by the same vendor or integrated through custom middleware. That is not an internet. That is a collection of intranets.

I think the hardest problem in multiagent systems is not making each agent more intelligent. It is designing the coordination architecture so that N agents reliably produce better outcomes than one agent working alone, and we have no theory for that yet. Baird and Maruping gave us the dyadic starting point. Stelmaszak showed us the direction is bidirectional. Grisold showed us norms can help. But a full theory of multiagent coordination would need to explain how appraisal scales to networks, how distribution works without a central allocator, and how coordination emerges when no single actor manages the interdependencies. I wrote about delegation replacing use for single-agent systems in a previous post, and the logic there was that IS theory needed a construct update. For multiagent systems, the construct update is even more fundamental. We are not just replacing use with delegation. We are adding a fourth mechanism that Baird and Maruping did not name: arbitration, the process by which an ecosystem of agents with different preference functions resolves conflicts and produces a coherent outcome.

Organizations that are building multiagent systems today should pay less attention to the reasoning capabilities of individual agents and more attention to the coordination architecture that connects them. The agent that makes the best decision in isolation can be the agent that breaks the system when it acts without regard for its peers. The preference functions matter more than the inference speed. The governance rules matter more than the prompt templates. And the protocol for task handoff matters more than the model size. Because in a multiagent system, the weakest link is never a single agent. It is the space between them.