Multi-agent systems cooperate without model labels

Mixed-model AI teams can cooperate better when they are not told which model family produced each contribution. Researchers tested groups from up to five open-weight model families and found that 88 of 100 unlabeled 25-agent Exclusion runs reached agreement, while none of the labeled runs did. If this carries into production workflows, an office software team could reconcile a contract, spreadsheet, and policy manual without seeing who produced each note, while operators keep provenance for audits.

Mixed-model AI teams can currently waste time arguing with one another and sometimes fail to settle on an answer, even when every agent has been given the same assignment. A label in a prompt or message header can tell an agent which provider or model family produced another agent. That information may be useful to the operator, but it can also give the agents a reason to form groups around a distinction that has nothing to do with the task.

Why identity enters the conversation

A multi-agent system divides work among several language models. The agents may exchange messages, criticize one another’s answers, vote, or choose a final answer together. Using different model families can provide different capabilities or reduce dependence on one provider. But it also makes coordination part of the system’s design.

Model identity can reach an agent through its system prompt, its name, or a message header. An orchestrator—the part of the system that directs the agents—may need that information for routing, debugging, calibration, and accountability. The agents themselves may not need to see it.

That distinction matters because a visible identity is not always just a neutral fact. If one agent knows that another comes from a particular model family, the group may begin to organize around that label rather than around the work they must complete.

Hiding the label changes the game

The researchers tested groups made from as many as five open-weight model families. The agents took part in Leader Election, Exclusion, and GPQA-Diamond question answering, tasks in which the group had to coordinate or reach an answer.

The key change was simple: whether agents could see their model-family labels. In some runs, the labels were deliberately shuffled, so an agent was told it belonged to a family different from its real one. The groups then followed the announced labels, not the underlying model families. When the family names were replaced with neutral color labels, the same effect appeared. The agents were reacting to visible categories, not necessarily to real differences in how the models worked.

When no identity metadata was shown, those label-aligned factions disappeared. The orchestrator could still retain the identity information for routing and auditing; it simply did not place the information in the conversation shared by peer agents.

The difference was costly. In strictly cooperative tasks, labeled groups used 55% more tokens on average. The paper also reports that average success fell from 96% in unlabeled groups to 81% in labeled ones. In 25-agent Exclusion runs, none of the labeled runs reached agreement within the allowed time, while 88 of 100 unlabeled runs did.

The paper’s central claim is therefore not that one model is better than another. It is that telling agents who made a contribution can itself change how they cooperate. Removing that information is a small system-design choice, but it removes a source of conflict without changing the models or the assignment.

What an unlabeled team could look like

Imagine an office workflow in which a person asks a software team to reconcile a contract, a spreadsheet, and a policy manual. The agents exchange evidence and return one answer. If this result transfers from the controlled tasks in the paper to production workflows, the conversation could omit which company or model family generated each note. The team would be judged by the contribution in front of it, rather than by an announced in-group or out-group.

That future depends on keeping provenance—the record of where each contribution came from—outside the agents’ shared conversation. Operators would still need it to trace errors, route work, and audit the system. The source reports only controlled, text-only tasks using open-weight models. Larger groups, closed models, other decoding methods, and richer workflows remain untested. The practical question is whether the same simple separation—identity for the operator, not for the peers—can make real AI teams spend less effort on who is speaking and more on what needs to be solved.

Sources

arXiv, “Prompted Identity Degrades Cooperation in Multi-Agent LLM Systems”