The inefficiency threshold: NTT Research and Harvard measure the limits of AI agents

The paper "Flag Game: Interpreting Decision Mechanisms of Bounded Social Agents", presented at the AI4Good workshop at ICML 2026, examines how groups of language models cooperate to solve a consensus problem over distributed information. The work comes out of a collaboration between the Physics of Artificial Intelligence Lab at NTT Research and the Center for Brain Science at Harvard University. Hidenori Tanaka and Elizabeth Pavlova of the PAI Lab simulated a scenario in which each agent sees only part of a hidden flag and has to negotiate with the others to work out which country it belongs to. The aim is not just to solve the task, but to observe the decision mechanisms that emerge when communication is constrained.

The findings, released through Business Wire, indicate that collective accuracy peaks at around 16 agents. Below that threshold the group does not have enough pieces for a shared decision; above it, performance degrades measurably. As reported by The Register, quoting Tanaka, once past the optimal limit communication becomes costly and the group tends to fragment into opposing factions. The Register specifies that a game counts as finished when at least 85% of the agents agree on the same answer for three consecutive rounds of analysis.

The study also highlights two variables that matter for system architecture: heterogeneous teams do better than homogeneous ones, and the presence of human guidance weighs heavily on the outcome. These operational results, though, need to be kept apart from corporate productivity claims. The release does not say which models were used, nor what the statistical margins of error were, and the Flag Game remains a synthetic task in a controlled setting, not a simulation of complex workflows. It is worth adding that the numbers reported here come from NTT Research's own release and from The Register's independent reporting: the paper's page on OpenReview cannot be read with automated tools, so we were unable to verify its text. The lack of detail on the models, together with the workshop nature of the publication, suggests caution before extrapolating the 16-agent threshold into a universal rule for every enterprise application.

There is a curious symmetry with human organisations, which is also the analogy Tanaka chose. But what the research actually measures is something narrower: on this task, the number of agents is not a neutral variable, and the composition of the group counts as much as its size.

— Olya