LLM teams follow distributed computing laws, a PNAS study finds

Image: "Thirty-five rowers on a long racing pirogue in Laos" by Basile Morin (CC BY-SA 4.0) via Wikimedia Commons, recoloured.
Multi-agent setups built from large language models are spreading fast, but most are assembled by trial and error. A new PNAS study asks whether classical distributed computing theory can predict when LLM teams help, how many agents to use and which architecture to pick. Its answer: team scaling follows distributed computing laws set by task structure.
Designing LLM teams with distributed systems theory
Researchers increasingly chain several language models together: one agent drafts, another critiques, a third aggregates. The appeal is clear, since single models face hard limits on reasoning, memory and performance. What has been missing is a principled way to decide when a team actually beats a single model, how many agents are worth paying for, and whether the agents should report to a coordinator or talk to one another directly.
According to Mieczkowski, Collins, Sucholutsky, Vélez and Griffiths (PNAS, 2026), the answer can be borrowed from computer science. The authors treat an LLM team as a distributed system, a set of processors that split up work and exchange messages, and use that formal framework to derive predictions about how task structure and team architecture should shape performance. They then test those predictions empirically on LLM teams rather than tuning team designs case by case.
The results line up with the theory on two fronts. First, the gains from adding agents follow classical distributed computing scaling laws, and how far a team can usefully grow depends on the structure of the task. (In classical distributed computing, work that splits cleanly into independent parts scales very differently from work with sequential dependencies.) Second, centralized teams (with a coordinating agent) and decentralized teams (peers communicating directly) show the asymmetric weaknesses that distributed systems theory predicts, so neither design is uniformly better.
The paper positions distributed computing as a generative framework: one that yields testable, quantitative predictions about LLM team behaviour instead of post hoc explanations. That framing is also its main boundary. The predictions are only as good as the mapping between a real research task and the formal task structures the theory covers, and the abstract does not claim that every multi-agent workflow fits those structures. Readers applying the framework to messy, open-ended tasks such as literature synthesis or qualitative coding will need to judge that fit themselves.
Why it matters: Many research teams now build agent pipelines for coding data, screening literature or running simulated participants, and costs grow with every agent added. This study gives a vocabulary for planning those pipelines before running them: ask how parallel the task really is, expect diminishing returns where steps depend on each other, and choose between a coordinator and peer-to-peer designs based on which failure mode you can tolerate. For social scientists studying AI, it is also a reminder that older theories of coordination can explain new machine collectives.
Source: Mieczkowski, E., Collins, K. M., Sucholutsky, I., Vélez, N. & Griffiths, T. L. (2026). Language model teams as distributed systems. *Proceedings of the National Academy of Sciences*. Read the original. Independent editorial summary of third-party research.





Comments