When More Capable AI Agents Make a Worse Team

Picture three research agents reading the same stale dashboard at 9 a.m., each reporting the same confident number as if it were fresh. Add a fourth, smarter agent and the team doesn't get sharper; it gets louder. The mistake isn't in any single model. It is in how the collective built its picture of the world.

A new preprint on multi-agent AI systems, titled "When Do Institutions Beat Intelligence?", argues that more capable agents do not necessarily form a more capable collective. The point is not that smarter models are useless. The point is that wiring them into a team is a separate design problem, and the rules a collective uses to route, admit, maintain, and act on shared information (what the paper calls institutions) can either repair a broken collective or get in the way of one that already works.

The authors build controlled "artificial ecologies" around four loci where multi-agent systems tend to fail. Naming them turns hand-waving about "agent orchestration" into a diagnostic. The first is access and routing: evidence never reaches the right decision-maker, or the right evidence is filtered out before it can be checked. The second is admission and dependence: reports that look independent trace back to one upstream signal, so a single error looks like several confirming sources. The third is state maintenance and incentives: the shared picture of the world goes stale, or gets shaped by what agents are rewarded to report rather than what is true. The fourth is representation and action: useful evidence is exposed through an interface the collective cannot reliably act on, so the right answer is never the one that gets executed.

These are not exotic failure modes. Anyone who has watched a dashboard-driven team double down on a number that was already outdated by lunchtime, or seen three analysts cite the same primary source as if it were three sources, has watched a version of this in the slow world. The paper gives the failure modes names, and a test for which one a team is looking at.

Institutions help when they repair a failure in how a collective constructs usable public state. They lose their advantage in three cases: when their signals are uninformative or uncheckable, when a more capable model can perform the same translation on its own, and when the resulting state still cannot drive a reliable action. That recasts the choice between "add a smarter model" and "add a rule" as a question of which of the four loci is actually broken.

Multi-agent AI systems are no longer a research curiosity. They are being wired into research, coding, and operations pipelines inside companies that already had trouble keeping one model honest. The paper does not tell builders which institutional design to pick. It tells them which failure they are looking at, and warns that stronger intelligence is sometimes the right substitute for a new process, the opposite of the default instinct in both vendor pitches and internal governance reviews.

The framework is preprint, not peer-reviewed, and the authors stress that the result comes from artificial ecologies, not live organizations. The paper's last move is a question rather than a prescription: before adding a model or a rule, name the failure mode. Once the failure is named, the choice between a smarter model and a new rule usually makes itself.